-
Notifications
You must be signed in to change notification settings - Fork 58
/
chap1.html
2215 lines (2148 loc) · 187 KB
/
chap1.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
<!DOCTYPE html>
<html lang="en">
<!-- Produced from a LaTeX source file. Note that the production is done -->
<!-- by a very rough-and-ready (and buggy) script, so the HTML and other -->
<!-- code is quite ugly! Later versions should be better. -->
<head>
<meta charset="utf-8">
<meta name="citation_title" content="Neural Networks and Deep Learning">
<meta name="citation_author" content="Nielsen, Michael A.">
<meta name="citation_publication_date" content="2015">
<meta name="citation_fulltext_html_url" content="http://neuralnetworksanddeeplearning.com">
<meta name="citation_publisher" content="Determination Press">
<meta name="citation_fulltext_world_readable" content="">
<link rel="icon" href="nnadl_favicon.ICO" />
<title>Neural networks and deep learning</title>
<script src="assets/jquery.min.js"></script>
<script type="text/x-mathjax-config">
MathJax.Hub.Config({
tex2jax: {inlineMath: [['$','$']]},
"HTML-CSS":
{scale: 92},
TeX: { equationNumbers: { autoNumber: "AMS" }}});
</script>
<script type="text/javascript" src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
<link href="assets/style.css" rel="stylesheet">
<link href="assets/pygments.css" rel="stylesheet">
<link rel="stylesheet" href="https://code.jquery.com/ui/1.11.2/themes/smoothness/jquery-ui.css">
<style>
/* Adapted from */
/* https://groups.google.com/d/msg/mathjax-users/jqQxrmeG48o/oAaivLgLN90J, */
/* by David Cervone */
@font-face {
font-family: 'MJX_Math';
src: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/eot/MathJax_Math-Italic.eot'); /* IE9 Compat Modes */
src: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/eot/MathJax_Math-Italic.eot?iefix') format('eot'),
url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/woff/MathJax_Math-Italic.woff') format('woff'),
url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/otf/MathJax_Math-Italic.otf') format('opentype'),
url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/svg/MathJax_Math-Italic.svg#MathJax_Math-Italic') format('svg');
}
@font-face {
font-family: 'MJX_Main';
src: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/eot/MathJax_Main-Regular.eot'); /* IE9 Compat Modes */
src: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/eot/MathJax_Main-Regular.eot?iefix') format('eot'),
url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/woff/MathJax_Main-Regular.woff') format('woff'),
url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/otf/MathJax_Main-Regular.otf') format('opentype'),
url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/fonts/HTML-CSS/TeX/svg/MathJax_Main-Regular.svg#MathJax_Main-Regular') format('svg');
}
</style>
</head>
<body><div class="header"><h1 class="chapter_number">
<a href="">CHAPTER 1</a></h1>
<h1 class="chapter_title"><a href="">Using neural nets to recognize handwritten digits</a></h1></div><div class="section"><div id="toc">
<p class="toc_title"><a href="index.html">Neural Networks and Deep Learning</a></p><p class="toc_not_mainchapter"><a href="about.html">What this book is about</a></p><p class="toc_not_mainchapter"><a href="exercises_and_problems.html">On the exercises and problems</a></p><p class='toc_mainchapter'><a id="toc_using_neural_nets_to_recognize_handwritten_digits_reveal" class="toc_reveal" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';"><img id="toc_img_using_neural_nets_to_recognize_handwritten_digits" src="images/arrow.png" width="15px"></a><a href="chap1.html">Using neural nets to recognize handwritten digits</a><div id="toc_using_neural_nets_to_recognize_handwritten_digits" style="display: none;"><p class="toc_section"><ul><a href="chap1.html#perceptrons"><li>Perceptrons</li></a><a href="chap1.html#sigmoid_neurons"><li>Sigmoid neurons</li></a><a href="chap1.html#the_architecture_of_neural_networks"><li>The architecture of neural networks</li></a><a href="chap1.html#a_simple_network_to_classify_handwritten_digits"><li>A simple network to classify handwritten digits</li></a><a href="chap1.html#learning_with_gradient_descent"><li>Learning with gradient descent</li></a><a href="chap1.html#implementing_our_network_to_classify_digits"><li>Implementing our network to classify digits</li></a><a href="chap1.html#toward_deep_learning"><li>Toward deep learning</li></a></ul></p></div>
<script>
$('#toc_using_neural_nets_to_recognize_handwritten_digits_reveal').click(function() {
var src = $('#toc_img_using_neural_nets_to_recognize_handwritten_digits').attr('src');
if(src == 'images/arrow.png') {
$("#toc_img_using_neural_nets_to_recognize_handwritten_digits").attr('src', 'images/arrow_down.png');
} else {
$("#toc_img_using_neural_nets_to_recognize_handwritten_digits").attr('src', 'images/arrow.png');
};
$('#toc_using_neural_nets_to_recognize_handwritten_digits').toggle('fast', function() {});
});</script><p class='toc_mainchapter'><a id="toc_how_the_backpropagation_algorithm_works_reveal" class="toc_reveal" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';"><img id="toc_img_how_the_backpropagation_algorithm_works" src="images/arrow.png" width="15px"></a><a href="chap2.html">How the backpropagation algorithm works</a><div id="toc_how_the_backpropagation_algorithm_works" style="display: none;"><p class="toc_section"><ul><a href="chap2.html#warm_up_a_fast_matrix-based_approach_to_computing_the_output_from_a_neural_network"><li>Warm up: a fast matrix-based approach to computing the output from a neural network</li></a><a href="chap2.html#the_two_assumptions_we_need_about_the_cost_function"><li>The two assumptions we need about the cost function</li></a><a href="chap2.html#the_hadamard_product_$s_\odot_t$"><li>The Hadamard product, $s \odot t$</li></a><a href="chap2.html#the_four_fundamental_equations_behind_backpropagation"><li>The four fundamental equations behind backpropagation</li></a><a href="chap2.html#proof_of_the_four_fundamental_equations_(optional)"><li>Proof of the four fundamental equations (optional)</li></a><a href="chap2.html#the_backpropagation_algorithm"><li>The backpropagation algorithm</li></a><a href="chap2.html#the_code_for_backpropagation"><li>The code for backpropagation</li></a><a href="chap2.html#in_what_sense_is_backpropagation_a_fast_algorithm"><li>In what sense is backpropagation a fast algorithm?</li></a><a href="chap2.html#backpropagation_the_big_picture"><li>Backpropagation: the big picture</li></a></ul></p></div>
<script>
$('#toc_how_the_backpropagation_algorithm_works_reveal').click(function() {
var src = $('#toc_img_how_the_backpropagation_algorithm_works').attr('src');
if(src == 'images/arrow.png') {
$("#toc_img_how_the_backpropagation_algorithm_works").attr('src', 'images/arrow_down.png');
} else {
$("#toc_img_how_the_backpropagation_algorithm_works").attr('src', 'images/arrow.png');
};
$('#toc_how_the_backpropagation_algorithm_works').toggle('fast', function() {});
});</script><p class='toc_mainchapter'><a id="toc_improving_the_way_neural_networks_learn_reveal" class="toc_reveal" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';"><img id="toc_img_improving_the_way_neural_networks_learn" src="images/arrow.png" width="15px"></a><a href="chap3.html">Improving the way neural networks learn</a><div id="toc_improving_the_way_neural_networks_learn" style="display: none;"><p class="toc_section"><ul><a href="chap3.html#the_cross-entropy_cost_function"><li>The cross-entropy cost function</li></a><a href="chap3.html#overfitting_and_regularization"><li>Overfitting and regularization</li></a><a href="chap3.html#weight_initialization"><li>Weight initialization</li></a><a href="chap3.html#handwriting_recognition_revisited_the_code"><li>Handwriting recognition revisited: the code</li></a><a href="chap3.html#how_to_choose_a_neural_network's_hyper-parameters"><li>How to choose a neural network's hyper-parameters?</li></a><a href="chap3.html#other_techniques"><li>Other techniques</li></a></ul></p></div>
<script>
$('#toc_improving_the_way_neural_networks_learn_reveal').click(function() {
var src = $('#toc_img_improving_the_way_neural_networks_learn').attr('src');
if(src == 'images/arrow.png') {
$("#toc_img_improving_the_way_neural_networks_learn").attr('src', 'images/arrow_down.png');
} else {
$("#toc_img_improving_the_way_neural_networks_learn").attr('src', 'images/arrow.png');
};
$('#toc_improving_the_way_neural_networks_learn').toggle('fast', function() {});
});</script><p class='toc_mainchapter'><a id="toc_a_visual_proof_that_neural_nets_can_compute_any_function_reveal" class="toc_reveal" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';"><img id="toc_img_a_visual_proof_that_neural_nets_can_compute_any_function" src="images/arrow.png" width="15px"></a><a href="chap4.html">A visual proof that neural nets can compute any function</a><div id="toc_a_visual_proof_that_neural_nets_can_compute_any_function" style="display: none;"><p class="toc_section"><ul><a href="chap4.html#two_caveats"><li>Two caveats</li></a><a href="chap4.html#universality_with_one_input_and_one_output"><li>Universality with one input and one output</li></a><a href="chap4.html#many_input_variables"><li>Many input variables</li></a><a href="chap4.html#extension_beyond_sigmoid_neurons"><li>Extension beyond sigmoid neurons</li></a><a href="chap4.html#fixing_up_the_step_functions"><li>Fixing up the step functions</li></a><a href="chap4.html#conclusion"><li>Conclusion</li></a></ul></p></div>
<script>
$('#toc_a_visual_proof_that_neural_nets_can_compute_any_function_reveal').click(function() {
var src = $('#toc_img_a_visual_proof_that_neural_nets_can_compute_any_function').attr('src');
if(src == 'images/arrow.png') {
$("#toc_img_a_visual_proof_that_neural_nets_can_compute_any_function").attr('src', 'images/arrow_down.png');
} else {
$("#toc_img_a_visual_proof_that_neural_nets_can_compute_any_function").attr('src', 'images/arrow.png');
};
$('#toc_a_visual_proof_that_neural_nets_can_compute_any_function').toggle('fast', function() {});
});</script><p class='toc_mainchapter'><a id="toc_why_are_deep_neural_networks_hard_to_train_reveal" class="toc_reveal" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';"><img id="toc_img_why_are_deep_neural_networks_hard_to_train" src="images/arrow.png" width="15px"></a><a href="chap5.html">Why are deep neural networks hard to train?</a><div id="toc_why_are_deep_neural_networks_hard_to_train" style="display: none;"><p class="toc_section"><ul><a href="chap5.html#the_vanishing_gradient_problem"><li>The vanishing gradient problem</li></a><a href="chap5.html#what's_causing_the_vanishing_gradient_problem_unstable_gradients_in_deep_neural_nets"><li>What's causing the vanishing gradient problem? Unstable gradients in deep neural nets</li></a><a href="chap5.html#unstable_gradients_in_more_complex_networks"><li>Unstable gradients in more complex networks</li></a><a href="chap5.html#other_obstacles_to_deep_learning"><li>Other obstacles to deep learning</li></a></ul></p></div>
<script>
$('#toc_why_are_deep_neural_networks_hard_to_train_reveal').click(function() {
var src = $('#toc_img_why_are_deep_neural_networks_hard_to_train').attr('src');
if(src == 'images/arrow.png') {
$("#toc_img_why_are_deep_neural_networks_hard_to_train").attr('src', 'images/arrow_down.png');
} else {
$("#toc_img_why_are_deep_neural_networks_hard_to_train").attr('src', 'images/arrow.png');
};
$('#toc_why_are_deep_neural_networks_hard_to_train').toggle('fast', function() {});
});</script><p class='toc_mainchapter'><a id="toc_deep_learning_reveal" class="toc_reveal" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';"><img id="toc_img_deep_learning" src="images/arrow.png" width="15px"></a><a href="chap6.html">Deep learning</a><div id="toc_deep_learning" style="display: none;"><p class="toc_section"><ul><a href="chap6.html#introducing_convolutional_networks"><li>Introducing convolutional networks</li></a><a href="chap6.html#convolutional_neural_networks_in_practice"><li>Convolutional neural networks in practice</li></a><a href="chap6.html#the_code_for_our_convolutional_networks"><li>The code for our convolutional networks</li></a><a href="chap6.html#recent_progress_in_image_recognition"><li>Recent progress in image recognition</li></a><a href="chap6.html#other_approaches_to_deep_neural_nets"><li>Other approaches to deep neural nets</li></a><a href="chap6.html#on_the_future_of_neural_networks"><li>On the future of neural networks</li></a></ul></p></div>
<script>
$('#toc_deep_learning_reveal').click(function() {
var src = $('#toc_img_deep_learning').attr('src');
if(src == 'images/arrow.png') {
$("#toc_img_deep_learning").attr('src', 'images/arrow_down.png');
} else {
$("#toc_img_deep_learning").attr('src', 'images/arrow.png');
};
$('#toc_deep_learning').toggle('fast', function() {});
});</script><p class="toc_not_mainchapter"><a href="sai.html">Appendix: Is there a <em>simple</em> algorithm for intelligence?</a></p><p class="toc_not_mainchapter"><a href="acknowledgements.html">Acknowledgements</a></p><p class="toc_not_mainchapter"><a href="faq.html">Frequently Asked Questions</a></p>
<hr>
<p class="sidebar"> If you benefit from the book, please make a small
donation. I suggest $5, but you can choose the amount.</p>
<form action="https://www.paypal.com/cgi-bin/webscr" method="post" target="_top">
<input type="hidden" name="cmd" value="_s-xclick">
<input type="hidden" name="hosted_button_id" value="5K9YAHR4X84RN">
<input type="image" src="https://www.paypalobjects.com/en_US/i/btn/btn_donateCC_LG.gif" border="0" name="submit" alt="PayPal - The safer, easier way to pay online!">
<img alt="" border="0" src="https://www.paypalobjects.com/en_US/i/scr/pixel.gif" width="1" height="1">
</form>
<p class="sidebar">Alternately, you can make a donation by sending me
Bitcoin, at address <span style="font-size: 0.7em">1Kd6tXH5SDAmiFb49J9hknG5pqj7KStSAx</span></p>
<!--
<hr>
<p class="sidebar"> If you benefit from the book, please make a small
donation. I suggest $3, but you can choose the amount.</p>
<form action="https://www.paypal.com/cgi-bin/webscr" method="post" target="_top">
<input type="hidden" name="cmd" value="_s-xclick">
<input type="hidden" name="encrypted" value="-----BEGIN PKCS7-----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-----END PKCS7-----
">
<input type="image" src="https://www.paypalobjects.com/en_US/i/btn/btn_donateCC_LG.gif" border="0" name="submit" alt="PayPal - The safer, easier way to pay online!">
<img alt="" border="0" src="https://www.paypalobjects.com/en_US/i/scr/pixel.gif" width="1" height="1">
</form>
-->
<hr>
<span class="sidebar_title">Sponsors</span>
<br/>
<a href="https://lambdalabs.com/?utm_source=neuralnetworksdeeplearning&utm_medium=banner&utm_campaign=blogin&utm_content=rbannerimg">
<img src="assets/lambda.png" width="200px" style="padding: 3px 0px 0px 10px; border-style: none;">
</a>
<br>
<div style="line-height: 1.2; padding-bottom: 12px; font-size: 0.8;">
<a href="https://lambdalabs.com/?utm_source=neuralnetworksdeeplearning&utm_medium=banner&utm_campaign=blogin&utm_content=rtext">Deep Learning Workstations, Servers, and Laptops</a>
</div>
<a href='http://gsquaredcapital.com/'><img src='assets/gsquared.png' width='200px' style="padding: 5px 0px 10px 10px; border-style: none;"></a>
<a href='http://www.tineye.com'><img src='assets/tineye.png' width='200px'
style="padding: 0px 0px 10px 8px; border-style: none;"></a>
<a href='http://www.visionsmarts.com'><img
src='assets/visionsmarts.png' width='210px' style="padding: 0px 0px
0px 0px; border-style: none;"></a> <br/>
<p class="sidebar">Thanks to all the <a
href="supporters.html">supporters</a> who made the book possible, with
especial thanks to Pavel Dudrenov. Thanks also to all the
contributors to the <a href="bugfinder.html">Bugfinder Hall of
Fame</a>. </p>
<hr>
<span class="sidebar_title">Resources</span>
<p class="sidebar"><a href="https://twitter.com/michael_nielsen">Michael Nielsen on Twitter</a></p>
<p class="sidebar"><a href="faq.html">Book FAQ</a></p>
<p class="sidebar">
<a href="https://github.com/mnielsen/neural-networks-and-deep-learning">Code repository</a></p>
<p class="sidebar">
<a href="http://eepurl.com/0Xxjb">Michael Nielsen's project announcement mailing list</a>
</p>
<p class="sidebar"> <a href="http://www.deeplearningbook.org/">Deep Learning</a>, book by Ian
Goodfellow, Yoshua Bengio, and Aaron Courville</p>
<p class="sidebar"><a href="http://cognitivemedium.com">cognitivemedium.com</a></p>
<hr>
<a href="http://michaelnielsen.org"><img src="assets/Michael_Nielsen_Web_Small.jpg" width="160px" style="border-style: none;"/></a>
<p class="sidebar">
By <a href="http://michaelnielsen.org">Michael Nielsen</a> / Dec 2019
</p>
</div>
</p><p>The human visual system is one of the wonders of the world. Considerthe following sequence of handwritten digits: <a name="complete_zero"></a></p><p><center><img src="images/digits.png" width="160px"></center> </p><p>Most people effortlessly recognize those digits as 504192. That easeis deceptive. In each hemisphere of our brain, humans have a primaryvisual cortex, also known as V1, containing 140 million neurons, withtens of billions of connections between them. And yet human visioninvolves not just V1, but an entire series of visual cortices - V2,V3, V4, and V5 - doing progressively more complex image processing.We carry in our heads a supercomputer, tuned by evolution overhundreds of millions of years, and superbly adapted to understand thevisual world. Recognizing handwritten digits isn't easy. Rather, wehumans are stupendously, astoundingly good at making sense of what oureyes show us. But nearly all that work is done unconsciously. And sowe don't usually appreciate how tough a problem our visual systemssolve.</p><p>The difficulty of visual pattern recognition becomes apparent if youattempt to write a computer program to recognize digits like thoseabove. What seems easy when we do it ourselves suddenly becomesextremely difficult. Simple intuitions about how we recognize shapes- "a 9 has a loop at the top, and a vertical stroke in the bottomright" - turn out to be not so simple to express algorithmically.When you try to make such rules precise, you quickly get lost in amorass of exceptions and caveats and special cases. It seemshopeless.</p><p></p><p>Neural networks approach the problem in a different way. The idea isto take a large number of handwritten digits, known as trainingexamples,</p><p><center><img src="images/mnist_100_digits.png" width="440px"></center></p><p>and then develop a system which can learn from those trainingexamples. In other words, the neural network uses the examples toautomatically infer rules for recognizing handwritten digits.Furthermore, by increasing the number of training examples, thenetwork can learn more about handwriting, and so improve its accuracy.So while I've shown just 100 training digits above, perhaps we couldbuild a better handwriting recognizer by using thousands or evenmillions or billions of training examples.</p><p>In this chapter we'll write a computer program implementing a neuralnetwork that learns to recognize handwritten digits. The program isjust 74 lines long, and uses no special neural network libraries. Butthis short program can recognize digits with an accuracy over 96percent, without human intervention. Furthermore, in later chapterswe'll develop ideas which can improve accuracy to over 99 percent. Infact, the best commercial neural networks are now so good that theyare used by banks to process cheques, and by post offices to recognizeaddresses.</p><p>We're focusing on handwriting recognition because it's an excellentprototype problem for learning about neural networks in general. As aprototype it hits a sweet spot: it's challenging - it's no smallfeat to recognize handwritten digits - but it's not so difficult asto require an extremely complicated solution, or tremendouscomputational power. Furthermore, it's a great way to develop moreadvanced techniques, such as deep learning. And so throughout thebook we'll return repeatedly to the problem of handwritingrecognition. Later in the book, we'll discuss how these ideas may beapplied to other problems in computer vision, and also in speech,natural language processing, and other domains.</p><p>Of course, if the point of the chapter was only to write a computerprogram to recognize handwritten digits, then the chapter would bemuch shorter! But along the way we'll develop many key ideas aboutneural networks, including two important types of artificial neuron(the perceptron and the sigmoid neuron), and the standard learningalgorithm for neural networks, known as stochastic gradient descent.Throughout, I focus on explaining <em>why</em> things are done the waythey are, and on building your neural networks intuition. Thatrequires a lengthier discussion than if I just presented the basicmechanics of what's going on, but it's worth it for the deeperunderstanding you'll attain. Amongst the payoffs, by the end of thechapter we'll be in position to understand what deep learning is, andwhy it matters.</p><p><h3><a name="perceptrons"></a><a href="#perceptrons">Perceptrons</a></h3></p><p>What is a neural network? To get started, I'll explain a type ofartificial neuron called a <em>perceptron</em>.Perceptrons were<a href="http://books.google.ca/books/about/Principles_of_neurodynamics.html?id=7FhRAAAAMAAJ">developed</a>in the 1950s and 1960s by the scientist<a href="http://en.wikipedia.org/wiki/Frank_Rosenblatt">Frank Rosenblatt</a>, inspired by earlier<a href="http://scholar.google.ca/scholar?cluster=4035975255085082870">work</a>by <a href="http://en.wikipedia.org/wiki/Warren_McCulloch">Warren McCulloch</a> and<a href="http://en.wikipedia.org/wiki/Walter_Pitts">Walter Pitts</a>. Today, it's more common to use othermodels of artificial neurons - in this book, and in much modern workon neural networks, the main neuron model used is one called the<em>sigmoid neuron</em>. We'll get to sigmoid neurons shortly. But tounderstand why sigmoid neurons are defined the way they are, it'sworth taking the time to first understand perceptrons.</p><p>So how do perceptrons work? A perceptron takes several binary inputs,$x_1, x_2, \ldots$, and produces a single binary output:<center><img src="images/tikz0.png"/></center>In the example shown the perceptron has three inputs, $x_1, x_2, x_3$.In general it could have more or fewer inputs. Rosenblatt proposed asimple rule to compute the output. He introduced<em>weights</em>, $w_1,w_2,\ldots$, real numbersexpressing the importance of the respective inputs to the output. Theneuron's output, $0$ or $1$, is determined by whether the weighted sum$\sum_j w_j x_j$ is less than or greater than some <em>threshold value</em>. Just like the weights, thethreshold is a real number which is a parameter of the neuron. To putit in more precise algebraic terms:<a class="displaced_anchor" name="eqtn1"></a>\begin{eqnarray} \mbox{output} & = & \left\{ \begin{array}{ll} 0 & \mbox{if } \sum_j w_j x_j \leq \mbox{ threshold} \\ 1 & \mbox{if } \sum_j w_j x_j > \mbox{ threshold} \end{array} \right.\tag{1}\end{eqnarray}That's all there is to how a perceptron works!</p><p>That's the basic mathematical model. A way you can think about theperceptron is that it's a device that makes decisions by weighing upevidence. Let me give an example. It's not a very realistic example,but it's easy to understand, and we'll soon get to more realisticexamples. Suppose the weekend is coming up, and you've heard thatthere's going to be a cheese festival in your city. You like cheese,and are trying to decide whether or not to go to the festival. Youmight make your decision by weighing up three factors:<ol><li> Is the weather good?<li> Does your boyfriend or girlfriend want to accompany you?<li> Is the festival near public transit? (You don't own a car).</ol>We can represent these three factors by corresponding binary variables$x_1, x_2$, and $x_3$. For instance, we'd have $x_1 = 1$ if theweather is good, and $x_1 = 0$ if the weather is bad. Similarly, $x_2= 1$ if your boyfriend or girlfriend wants to go, and $x_2 = 0$ ifnot. And similarly again for $x_3$ and public transit.</p><p>Now, suppose you absolutely adore cheese, so much so that you're happyto go to the festival even if your boyfriend or girlfriend isuninterested and the festival is hard to get to. But perhaps youreally loathe bad weather, and there's no way you'd go to the festivalif the weather is bad. You can use perceptrons to model this kind ofdecision-making. One way to do this is to choose a weight $w_1 = 6$for the weather, and $w_2 = 2$ and $w_3 = 2$ for the other conditions.The larger value of $w_1$ indicates that the weather matters a lot toyou, much more than whether your boyfriend or girlfriend joins you, orthe nearness of public transit. Finally, suppose you choose athreshold of $5$ for the perceptron. With these choices, theperceptron implements the desired decision-making model, outputting$1$ whenever the weather is good, and $0$ whenever the weather is bad.It makes no difference to the output whether your boyfriend orgirlfriend wants to go, or whether public transit is nearby.</p><p>By varying the weights and the threshold, we can get different modelsof decision-making. For example, suppose we instead chose a thresholdof $3$. Then the perceptron would decide that you should go to thefestival whenever the weather was good <em>or</em> when both thefestival was near public transit <em>and</em> your boyfriend orgirlfriend was willing to join you. In other words, it'd be adifferent model of decision-making. Dropping the threshold meansyou're more willing to go to the festival.</p><p>Obviously, the perceptron isn't a complete model of humandecision-making! But what the example illustrates is how a perceptroncan weigh up different kinds of evidence in order to make decisions.And it should seem plausible that a complex network of perceptronscould make quite subtle decisions:<center><img src="images/tikz1.png"/></center>In this network, the first column of perceptrons - what we'll callthe first <em>layer</em> of perceptrons - is making three very simpledecisions, by weighing the input evidence. What about the perceptronsin the second layer? Each of those perceptrons is making a decisionby weighing up the results from the first layer of decision-making.In this way a perceptron in the second layer can make a decision at amore complex and more abstract level than perceptrons in the firstlayer. And even more complex decisions can be made by the perceptronin the third layer. In this way, a many-layer network of perceptronscan engage in sophisticated decision making.</p><p>Incidentally, when I defined perceptrons I said that a perceptron hasjust a single output. In the network above the perceptrons look likethey have multiple outputs. In fact, they're still single output.The multiple output arrows are merely a useful way of indicating thatthe output from a perceptron is being used as the input to severalother perceptrons. It's less unwieldy than drawing a single outputline which then splits.</p><p>Let's simplify the way we describe perceptrons. The condition $\sum_jw_j x_j > \mbox{threshold}$ is cumbersome, and we can make twonotational changes to simplify it. The first change is to write$\sum_j w_j x_j$ as a dot product, $w \cdot x \equiv \sum_j w_j x_j$,where $w$ and $x$ are vectors whose components are the weights andinputs, respectively. The second change is to move the threshold tothe other side of the inequality, and to replace it by what's known asthe perceptron's <em>bias</em>, $b \equiv-\mbox{threshold}$. Using the bias instead of the threshold, theperceptron rule can berewritten:<a class="displaced_anchor" name="eqtn2"></a>\begin{eqnarray} \mbox{output} = \left\{ \begin{array}{ll} 0 & \mbox{if } w\cdot x + b \leq 0 \\ 1 & \mbox{if } w\cdot x + b > 0 \end{array} \right.\tag{2}\end{eqnarray}You can think of the bias as a measure of how easy it is to get theperceptron to output a $1$. Or to put it in more biological terms,the bias is a measure of how easy it is to get the perceptron to<em>fire</em>. For a perceptron with a really big bias, it's extremelyeasy for the perceptron to output a $1$. But if the bias is verynegative, then it's difficult for the perceptron to output a $1$.Obviously, introducing the bias is only a small change in how wedescribe perceptrons, but we'll see later that it leads to furthernotational simplifications. Because of this, in the remainder of thebook we won't use the threshold, we'll always use the bias.</p><p>I've described perceptrons as a method for weighing evidence to makedecisions. Another way perceptrons can be used is to compute theelementary logical functions we usually think of as underlyingcomputation, functions such as <CODE>AND</CODE>, <CODE>OR</CODE>, and<CODE>NAND</CODE>. For example, suppose we have a perceptron with twoinputs, each with weight $-2$, and an overall bias of $3$. Here's ourperceptron:<center><img src="images/tikz2.png"/></center>Then we see that input $00$ produces output $1$, since$(-2)*0+(-2)*0+3 = 3$ is positive. Here, I've introduced the $*$symbol to make the multiplications explicit. Similar calculationsshow that the inputs $01$ and $10$ produce output $1$. But the input$11$ produces output $0$, since $(-2)*1+(-2)*1+3 = -1$ is negative.And so our perceptron implements a <CODE>NAND</CODE>gate!</p><p><a name="universality"></a></p><p>The <CODE>NAND</CODE> example shows that we can use perceptrons to computesimple logical functions. In fact, we can usenetworks of perceptrons to compute <em>any</em> logical function at all.The reason is that the <CODE>NAND</CODE> gate is universal forcomputation, that is, we can build any computation up out of<CODE>NAND</CODE> gates. For example, we can use <CODE>NAND</CODE> gates tobuild a circuit which adds two bits, $x_1$ and $x_2$. This requirescomputing the bitwise sum, $x_1 \oplus x_2$, as well as a carry bitwhich is set to $1$ when both $x_1$ and $x_2$ are $1$, i.e., the carrybit is just the bitwise product $x_1 x_2$:<center> <img src="images/tikz3.png"/></center>To get an equivalent network of perceptrons we replace all the<CODE>NAND</CODE> gates by perceptrons with two inputs, each with weight$-2$, and an overall bias of $3$. Here's the resulting network. Notethat I've moved the perceptron corresponding to the bottom right<CODE>NAND</CODE> gate a little, just to make it easier to draw the arrowson the diagram:<center> <img src="images/tikz4.png"/></center>One notable aspect of this network of perceptrons is that the outputfrom the leftmost perceptron is used twice as input to the bottommostperceptron. When I defined the perceptron model I didn't say whetherthis kind of double-output-to-the-same-place was allowed. Actually,it doesn't much matter. If we don't want to allow this kind of thing,then it's possible to simply merge the two lines, into a singleconnection with a weight of -4 instead of two connections with -2weights. (If you don't find this obvious, you should stop and proveto yourself that this is equivalent.) With that change, the networklooks as follows, with all unmarked weights equal to -2, all biasesequal to 3, and a single weight of -4, as marked:<center> <img src="images/tikz5.png"/></center>Up to now I've been drawing inputs like $x_1$ and $x_2$ as variablesfloating to the left of the network of perceptrons. In fact, it'sconventional to draw an extra layer of perceptrons - the <em>input layer</em> - to encode the inputs:<center> <img src="images/tikz6.png"/></center>This notation for input perceptrons, in which we have an output, butno inputs,<center><img src="images/tikz7.png"/></center>is a shorthand. It doesn't actually mean a perceptron with no inputs.To see this, suppose we did have a perceptron with no inputs. Thenthe weighted sum $\sum_j w_j x_j$ would always be zero, and so theperceptron would output $1$ if $b > 0$, and $0$ if $b \leq 0$. Thatis, the perceptron would simply output a fixed value, not the desiredvalue ($x_1$, in the example above). It's better to think of theinput perceptrons as not really being perceptrons at all, but ratherspecial units which are simply defined to output the desired values,$x_1, x_2,\ldots$.</p><p>The adder example demonstrates how a network of perceptrons can beused to simulate a circuit containing many <CODE>NAND</CODE> gates. Andbecause <CODE>NAND</CODE> gates are universal for computation, it followsthat perceptrons are also universal for computation.</p><p>The computational universality of perceptrons is simultaneouslyreassuring and disappointing. It's reassuring because it tells usthat networks of perceptrons can be as powerful as any other computingdevice. But it's also disappointing, because it makes it seem asthough perceptrons are merely a new type of <CODE>NAND</CODE> gate.That's hardly big news!</p><p>However, the situation is better than this view suggests. It turnsout that we can devise <em>learning algorithms</em> which canautomatically tune the weights and biases of a network of artificialneurons. This tuning happens in response to external stimuli, withoutdirect intervention by a programmer. These learning algorithms enableus to use artificial neurons in a way which is radically different toconventional logic gates. Instead of explicitly laying out a circuitof <CODE>NAND</CODE> and other gates, our neural networks can simply learnto solve problems, sometimes problems where it would be extremelydifficult to directly design a conventional circuit.</p><p><h3><a name="sigmoid_neurons"></a><a href="#sigmoid_neurons">Sigmoid neurons</a></h3></p><p>Learning algorithms sound terrific. But how can we devise suchalgorithms for a neural network? Suppose we have a network ofperceptrons that we'd like to use to learn to solve some problem. Forexample, the inputs to the network might be the raw pixel data from ascanned, handwritten image of a digit. And we'd like the network tolearn weights and biases so that the output from the network correctlyclassifies the digit. To see how learning might work, suppose we makea small change in some weight (or bias) in the network. What we'dlike is for this small change in weight to cause only a smallcorresponding change in the output from the network. As we'll see ina moment, this property will make learning possible. Schematically,here's what we want (obviously this network is too simple to dohandwriting recognition!):</p><p><center><img src="images/tikz8.png"/></center></p><p>If it were true that a small change in a weight (or bias) causes onlya small change in output, then we could use this fact to modify theweights and biases to get our network to behave more in the manner wewant. For example, suppose the network was mistakenly classifying animage as an "8" when it should be a "9". We could figure out howto make a small change in the weights and biases so the network gets alittle closer to classifying the image as a "9". And then we'drepeat this, changing the weights and biases over and over to producebetter and better output. The network would be learning.</p><p>The problem is that this isn't what happens when our network containsperceptrons. In fact, a small change in the weights or bias of anysingle perceptron in the network can sometimes cause the output ofthat perceptron to completely flip, say from $0$ to $1$. That flipmay then cause the behaviour of the rest of the network to completelychange in some very complicated way. So while your "9" might now beclassified correctly, the behaviour of the network on all the otherimages is likely to have completely changed in some hard-to-controlway. That makes it difficult to see how to gradually modify theweights and biases so that the network gets closer to the desiredbehaviour. Perhaps there's some clever way of getting around thisproblem. But it's not immediately obvious how we can get a network ofperceptrons to learn.</p><p>We can overcome this problem by introducing a new type of artificialneuron called a <em>sigmoid</em> neuron.Sigmoid neurons are similar to perceptrons, but modified so that smallchanges in their weights and bias cause only a small change in theiroutput. That's the crucial fact which will allow a network of sigmoidneurons to learn.</p><p>Okay, let me describe the sigmoid neuron. We'll depict sigmoidneurons in the same way we depicted perceptrons:<center><img src="images/tikz9.png"/></center>Just like a perceptron, the sigmoid neuron has inputs, $x_1, x_2,\ldots$. But instead of being just $0$ or $1$, these inputs can alsotake on any values <em>between</em> $0$ and $1$. So, for instance,$0.638\ldots$ is a valid input for a sigmoid neuron. Also just like aperceptron, the sigmoid neuron has weights for each input, $w_1, w_2,\ldots$, and an overall bias, $b$. But the output is not $0$ or $1$.Instead, it's $\sigma(w \cdot x+b)$, where $\sigma$ is called the<em>sigmoid function</em>*<span class="marginnote">*Incidentally, $\sigma$ is sometimes called the <em>logistic function</em>, and this new class of neurons called <em>logistic neurons</em>. It's useful to remember this terminology, since these terms are used by many people working with neural nets. However, we'll stick with the sigmoid terminology.</span>, and is definedby:<a class="displaced_anchor" name="eqtn3"></a>\begin{eqnarray} \sigma(z) \equiv \frac{1}{1+e^{-z}}.\tag{3}\end{eqnarray}To put it all a little more explicitly, the output of a sigmoid neuronwith inputs $x_1,x_2,\ldots$, weights $w_1,w_2,\ldots$, and bias $b$ is<a class="displaced_anchor" name="eqtn4"></a>\begin{eqnarray} \frac{1}{1+\exp(-\sum_j w_j x_j-b)}.\tag{4}\end{eqnarray}</p><p>At first sight, sigmoid neurons appear very different to perceptrons.The algebraic form of the sigmoid function may seem opaque andforbidding if you're not already familiar with it. In fact, there aremany similarities between perceptrons and sigmoid neurons, and thealgebraic form of the sigmoid function turns out to be more of atechnical detail than a true barrier to understanding.</p><p>To understand the similarity to the perceptron model, suppose $z\equiv w \cdot x + b$ is a large positive number. Then $e^{-z}\approx 0$ and so $\sigma(z) \approx 1$. In other words, when $z = w\cdot x+b$ is large and positive, the output from the sigmoid neuronis approximately $1$, just as it would have been for a perceptron.Suppose on the other hand that $z = w \cdot x+b$ is very negative.Then $e^{-z} \rightarrow \infty$, and $\sigma(z) \approx 0$. So when$z = w \cdot x +b$ is very negative, the behaviour of a sigmoid neuronalso closely approximates a perceptron. It's only when $w \cdot x+b$is of modest size that there's much deviation from the perceptronmodel.</p><p>What about the algebraic form of $\sigma$? How can we understandthat? In fact, the exact form of $\sigma$ isn't so important - whatreally matters is the shape of the function when plotted. Here's theshape:</p><p><div id="sigmoid_graph"><a name="sigmoid_graph"></a></div><script src="http://d3js.org/d3.v3.min.js"></script><script>function s(x) {return 1/(1+Math.exp(-x));}var m = [40, 120, 50, 120];var height = 290 - m[0] - m[2];var width = 600 - m[1] - m[3];var xmin = -5;var xmax = 5;var sample = 400;var x1 = d3.scale.linear().domain([0, sample]).range([xmin, xmax]);var data = d3.range(sample).map(function(d){ return { x: x1(d), y: s(x1(d))}; });var x = d3.scale.linear().domain([xmin, xmax]).range([0, width]);var y = d3.scale.linear() .domain([0, 1]) .range([height, 0]);var line = d3.svg.line() .x(function(d) { return x(d.x); }) .y(function(d) { return y(d.y); })var graph = d3.select("#sigmoid_graph") .append("svg") .attr("width", width + m[1] + m[3]) .attr("height", height + m[0] + m[2]) .append("g") .attr("transform", "translate(" + m[3] + "," + m[0] + ")");var xAxis = d3.svg.axis() .scale(x) .tickValues(d3.range(-4, 5, 1)) .orient("bottom")graph.append("g") .attr("class", "x axis") .attr("transform", "translate(0, " + height + ")") .call(xAxis);var yAxis = d3.svg.axis() .scale(y) .tickValues(d3.range(0, 1.01, 0.2)) .orient("left") .ticks(5)graph.append("g") .attr("class", "y axis") .call(yAxis);graph.append("path").attr("d", line(data));graph.append("text") .attr("class", "x label") .attr("text-anchor", "end") .attr("x", width/2) .attr("y", height+35) .text("z");graph.append("text") .attr("x", (width / 2)) .attr("y", -10) .attr("text-anchor", "middle") .style("font-size", "16px") .text("sigmoid function");</script></p><p>This shape is a smoothed out version of a step function:</p><p><div id="step_graph"></div><script>function s(x) {return x < 0 ? 0 : 1;}var m = [40, 120, 50, 120];var height = 290 - m[0] - m[2];var width = 600 - m[1] - m[3];var xmin = -5;var xmax = 5;var sample = 400;var x1 = d3.scale.linear().domain([0, sample]).range([xmin, xmax]);var data = d3.range(sample).map(function(d){ return { x: x1(d), y: s(x1(d))}; });var x = d3.scale.linear().domain([xmin, xmax]).range([0, width]);var y = d3.scale.linear() .domain([0,1]) .range([height, 0]);var line = d3.svg.line() .x(function(d) { return x(d.x); }) .y(function(d) { return y(d.y); })var graph = d3.select("#step_graph") .append("svg") .attr("width", width + m[1] + m[3]) .attr("height", height + m[0] + m[2]) .append("g") .attr("transform", "translate(" + m[3] + "," + m[0] + ")");var xAxis = d3.svg.axis() .scale(x) .tickValues(d3.range(-4, 5, 1)) .orient("bottom")graph.append("g") .attr("class", "x axis") .attr("transform", "translate(0, " + height + ")") .call(xAxis);var yAxis = d3.svg.axis() .scale(y) .tickValues(d3.range(0, 1.01, 0.2)) .orient("left") .ticks(5)graph.append("g") .attr("class", "y axis") .call(yAxis);graph.append("path").attr("d", line(data));graph.append("text") .attr("class", "x label") .attr("text-anchor", "end") .attr("x", width/2) .attr("y", height+35) .text("z");graph.append("text") .attr("x", (width / 2)) .attr("y", -10) .attr("text-anchor", "middle") .style("font-size", "16px") .text("step function");</script></p><p>If $\sigma$ had in fact been a step function, then the sigmoid neuronwould <em>be</em> a perceptron, since the output would be $1$ or $0$depending on whether $w\cdot x+b$ was positive ornegative*<span class="marginnote">*Actually, when $w \cdot x +b = 0$ the perceptron outputs $0$, while the step function outputs $1$. So, strictly speaking, we'd need to modify the step function at that one point. But you get the idea.</span>. By using the actual $\sigma$ function weget, as already implied above, a smoothed out perceptron. Indeed,it's the smoothness of the $\sigma$ function that is the crucial fact,not its detailed form. The smoothness of $\sigma$ means that smallchanges $\Delta w_j$ in the weights and $\Delta b$ in the bias willproduce a small change $\Delta \mbox{output}$ in the output from theneuron. In fact, calculus tells us that $\Delta \mbox{output}$ iswell approximated by<a class="displaced_anchor" name="eqtn5"></a>\begin{eqnarray} \Delta \mbox{output} \approx \sum_j \frac{\partial \, \mbox{output}}{\partial w_j} \Delta w_j + \frac{\partial \, \mbox{output}}{\partial b} \Delta b,\tag{5}\end{eqnarray}where the sum is over all the weights, $w_j$, and $\partial \,\mbox{output} / \partial w_j$ and $\partial \, \mbox{output} /\partialb$ denote partial derivatives of the $\mbox{output}$ with respect to$w_j$ and $b$, respectively. Don't panic if you're not comfortablewith partial derivatives! While the expression above lookscomplicated, with all the partial derivatives, it's actually sayingsomething very simple (and which is very good news): $\Delta\mbox{output}$ is a <em>linear function</em> of the changes $\Delta w_j$and $\Delta b$ in the weights and bias. This linearity makes it easyto choose small changes in the weights and biases to achieve anydesired small change in the output. So while sigmoid neurons havemuch of the same qualitative behaviour as perceptrons, they make itmuch easier to figure out how changing the weights and biases willchange the output.</p><p>If it's the shape of $\sigma$ which really matters, and not its exactform, then why use the particular form used for $\sigma$ inEquation <span id="margin_301539119283_reveal" class="equation_link">(3)</span><span id="margin_301539119283" class="marginequation" style="display: none;"><a href="chap1.html#eqtn3" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \sigma(z) \equiv \frac{1}{1+e^{-z}} \nonumber\end{eqnarray}</a></span><script>$('#margin_301539119283_reveal').click(function() {$('#margin_301539119283').toggle('slow', function() {});});</script>? In fact, later in the book we willoccasionally consider neurons where the output is $f(w \cdot x + b)$for some other <em>activation function</em> $f(\cdot)$. The main thingthat changes when we use a different activation function is that theparticular values for the partial derivatives inEquation <span id="margin_244684310360_reveal" class="equation_link">(5)</span><span id="margin_244684310360" class="marginequation" style="display: none;"><a href="chap1.html#eqtn5" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta \mbox{output} \approx \sum_j \frac{\partial \, \mbox{output}}{\partial w_j} \Delta w_j + \frac{\partial \, \mbox{output}}{\partial b} \Delta b \nonumber\end{eqnarray}</a></span><script>$('#margin_244684310360_reveal').click(function() {$('#margin_244684310360').toggle('slow', function() {});});</script> change. It turns out that when wecompute those partial derivatives later, using $\sigma$ will simplifythe algebra, simply because exponentials have lovely properties whendifferentiated. In any case, $\sigma$ is commonly-used in work onneural nets, and is the activation function we'll use most often inthis book.</p><p>How should we interpret the output from a sigmoid neuron? Obviously,one big difference between perceptrons and sigmoid neurons is thatsigmoid neurons don't just output $0$ or $1$. They can have as outputany real number between $0$ and $1$, so values such as $0.173\ldots$and $0.689\ldots$ are legitimate outputs. This can be useful, forexample, if we want to use the output value to represent the averageintensity of the pixels in an image input to a neural network. Butsometimes it can be a nuisance. Suppose we want the output from thenetwork to indicate either "the input image is a 9" or "the inputimage is not a 9". Obviously, it'd be easiest to do this if theoutput was a $0$ or a $1$, as in a perceptron. But in practice we canset up a convention to deal with this, for example, by deciding tointerpret any output of at least $0.5$ as indicating a "9", and anyoutput less than $0.5$ as indicating "not a 9". I'll alwaysexplicitly state when we're using such a convention, so it shouldn'tcause any confusion.</p><p><h4><a name="exercises_191892"></a><a href="#exercises_191892">Exercises</a></h4><ul><li><strong>Sigmoid neurons simulating perceptrons, part I</strong> $\mbox{}$ <br/> Suppose we take all the weights and biases in a network of perceptrons, and multiply them by a positive constant, $c > 0$. Show that the behaviour of the network doesn't change.</p><p><li><strong>Sigmoid neurons simulating perceptrons, part II</strong> $\mbox{}$ <br/> Suppose we have the same setup as the last problem - a network of perceptrons. Suppose also that the overall input to the network of perceptrons has been chosen. We won't need the actual input value, we just need the input to have been fixed. Suppose the weights and biases are such that $w \cdot x + b \neq 0$ for the input $x$ to any particular perceptron in the network. Now replace all the perceptrons in the network by sigmoid neurons, and multiply the weights and biases by a positive constant $c > 0$. Show that in the limit as $c \rightarrow \infty$ the behaviour of this network of sigmoid neurons is exactly the same as the network of perceptrons. How can this fail when $w \cdot x + b = 0$ for one of the perceptrons?</ul></p><p><h3><a name="the_architecture_of_neural_networks"></a><a href="#the_architecture_of_neural_networks">The architecture of neural networks</a></h3></p><p>In the next section I'll introduce a neural network that can do apretty good job classifying handwritten digits. In preparation forthat, it helps to explain some terminology that lets us name differentparts of a network. Suppose we have the network:<center><img src="images/tikz10.png"/></center>As mentioned earlier, the leftmost layer in this network is called theinput layer, and the neurons within thelayer are called <em>input neurons</em>.The rightmost or <em>output</em> layercontains the <em>output neurons</em>, or,as in this case, a single output neuron. The middle layer is called a<em>hidden layer</em>, since the neurons inthis layer are neither inputs nor outputs. The term "hidden"perhaps sounds a little mysterious - the first time I heard the termI thought it must have some deep philosophical or mathematicalsignificance - but it really means nothing more than "not an inputor an output". The network above has just a single hidden layer, butsome networks have multiple hidden layers. For example, the followingfour-layer network has two hidden layers:<center><img src="images/tikz11.png"/></center>Somewhat confusingly, and for historical reasons, such multiple layernetworks are sometimes called <em>multilayer perceptrons</em> or<em>MLPs</em>, despite being made up of sigmoid neurons, notperceptrons. I'm not going to use the MLP terminology in this book,since I think it's confusing, but wanted to warn you of its existence.</p><p>The design of the input and output layers in a network is oftenstraightforward. For example, suppose we're trying to determinewhether a handwritten image depicts a "9" or not. A natural way todesign the network is to encode the intensities of the image pixelsinto the input neurons. If the image is a $64$ by $64$ greyscaleimage, then we'd have $4,096 = 64 \times 64$ input neurons, with theintensities scaled appropriately between $0$ and $1$. The outputlayer will contain just a single neuron, with output values of lessthan $0.5$ indicating "input image is not a 9", and values greaterthan $0.5$ indicating "input image is a 9 ".</p><p></p><p></p><p>While the design of the input and output layers of a neural network isoften straightforward, there can be quite an art to the design of thehidden layers. In particular, it's not possible to sum up the designprocess for the hidden layers with a few simple rules of thumb.Instead, neural networks researchers have developed many designheuristics for the hidden layers, which help people get the behaviourthey want out of their nets. For example, such heuristics can be usedto help determine how to trade off the number of hidden layers againstthe time required to train the network. We'll meet several suchdesign heuristics later in this book. </p><p>Up to now, we've been discussing neural networks where the output fromone layer is used as input to the next layer. Such networks arecalled <em>feedforward</em>neural networks. This means there are no loops in the network -information is always fed forward, never fed back. If we did haveloops, we'd end up with situations where the input to the $\sigma$function depended on the output. That'd be hard to make sense of, andso we don't allow such loops.</p><p>However, there are other models of artificial neural networks in whichfeedback loops are possible. These models are called<a href="http://en.wikipedia.org/wiki/Recurrent_neural_network">recurrent neural networks</a>. The idea in these models is to have neurons whichfire for some limited duration of time, before becoming quiescent.That firing can stimulate other neurons, which may fire a little whilelater, also for a limited duration. That causes still more neurons tofire, and so over time we get a cascade of neurons firing. Loopsdon't cause problems in such a model, since a neuron's output onlyaffects its input at some later time, not instantaneously.</p><p></p><p>Recurrent neural nets have been less influential than feedforwardnetworks, in part because the learning algorithms for recurrent netsare (at least to date) less powerful. But recurrent networks arestill extremely interesting. They're much closer in spirit to how ourbrains work than feedforward networks. And it's possible thatrecurrent networks can solve important problems which can only besolved with great difficulty by feedforward networks. However, tolimit our scope, in this book we're going to concentrate on the morewidely-used feedforward networks.</p><p><h3><a name="a_simple_network_to_classify_handwritten_digits"></a><a href="#a_simple_network_to_classify_handwritten_digits">A simple network to classify handwritten digits</a></h3></p><p>Having defined neural networks, let's return to handwritingrecognition. We can split the problem of recognizing handwrittendigits into two sub-problems. First, we'd like a way of breaking animage containing many digits into a sequence of separate images, eachcontaining a single digit. For example, we'd like to break the image</p><p><center><img src="images/digits.png" width="300px"></center></p><p>into six separate images,</p><p><center><img src="images/digits_separate.png" width="440px"></center> </p><p>We humans solve this <em>segmentation problem</em> with ease, but it's challengingfor a computer program to correctly break up the image. Once theimage has been segmented, the program then needs to classify eachindividual digit. So, for instance, we'd like our program torecognize that the first digit above,</p><p><center><img src="images/mnist_first_digit.png" width="64px"></center></p><p>is a 5.</p><p>We'll focus on writing a program to solve the second problem, that is,classifying individual digits. We do this because it turns out thatthe segmentation problem is not so difficult to solve, once you have agood way of classifying individual digits. There are many approachesto solving the segmentation problem. One approach is to trial manydifferent ways of segmenting the image, using the individual digitclassifier to score each trial segmentation. A trial segmentationgets a high score if the individual digit classifier is confident ofits classification in all segments, and a low score if the classifieris having a lot of trouble in one or more segments. The idea is thatif the classifier is having trouble somewhere, then it's probablyhaving trouble because the segmentation has been chosen incorrectly.This idea and other variations can be used to solve the segmentationproblem quite well. So instead of worrying about segmentation we'llconcentrate on developing a neural network which can solve the moreinteresting and difficult problem, namely, recognizing individualhandwritten digits.</p><p>To recognize individual digits we will use a three-layer neuralnetwork:</p><p><center><img src="images/tikz12.png"/></center></p><p>The input layer of the network contains neurons encoding the values ofthe input pixels. As discussed in the next section, our training datafor the network will consist of many $28$ by $28$ pixel images ofscanned handwritten digits, and so the input layer contains $784 = 28\times 28$ neurons. For simplicity I've omitted most of the $784$input neurons in the diagram above. The input pixels are greyscale,with a value of $0.0$ representing white, a value of $1.0$representing black, and in between values representing graduallydarkening shades of grey.</p><p>The second layer of the network is a hidden layer. We denote thenumber of neurons in this hidden layer by $n$, and we'll experimentwith different values for $n$. The example shown illustrates a smallhidden layer, containing just $n = 15$ neurons.</p><p>The output layer of the network contains 10 neurons. If the firstneuron fires, i.e., has an output $\approx 1$, then that will indicatethat the network thinks the digit is a $0$. If the second neuronfires then that will indicate that the network thinks the digit is a$1$. And so on. A little more precisely, we number the outputneurons from $0$ through $9$, and figure out which neuron has thehighest activation value. If that neuron is, say, neuron number $6$,then our network will guess that the input digit was a $6$. And so onfor the other output neurons.</p><p>You might wonder why we use $10$ output neurons. After all, the goalof the network is to tell us which digit ($0, 1, 2, \ldots, 9$)corresponds to the input image. A seemingly natural way of doing thatis to use just $4$ output neurons, treating each neuron as taking on abinary value, depending on whether the neuron's output is closer to$0$ or to $1$. Four neurons are enough to encode the answer, since$2^4 = 16$ is more than the 10 possible values for the input digit.Why should our network use $10$ neurons instead? Isn't thatinefficient? The ultimate justification is empirical: we can try outboth network designs, and it turns out that, for this particularproblem, the network with $10$ output neurons learns to recognizedigits better than the network with $4$ output neurons. But thatleaves us wondering <em>why</em> using $10$ output neurons works better.Is there some heuristic that would tell us in advance that we shoulduse the $10$-output encoding instead of the $4$-output encoding?</p><p>To understand why we do this, it helps to think about what the neuralnetwork is doing from first principles. Consider first the case wherewe use $10$ output neurons. Let's concentrate on the first outputneuron, the one that's trying to decide whether or not the digit is a$0$. It does this by weighing up evidence from the hidden layer ofneurons. What are those hidden neurons doing? Well, just suppose forthe sake of argument that the first neuron in the hidden layer detectswhether or not an image like the following is present:</p><p><center><img src="images/mnist_top_left_feature.png" width="130px"></center></p><p>It can do this by heavily weighting input pixels which overlap withthe image, and only lightly weighting the other inputs. In a similarway, let's suppose for the sake of argument that the second, third,and fourth neurons in the hidden layer detect whether or not thefollowing images are present:</p><p><center><img src="images/mnist_other_features.png" width="424px"></center></p><p>As you may have guessed, these four images together make up the $0$image that we saw in the line of digits shown <a href="#complete_zero">earlier</a>:</p><p><center><img src="images/mnist_complete_zero.png" width="130px"></center></p><p>So if all four of these hidden neurons are firing then we can concludethat the digit is a $0$. Of course, that's not the <em>only</em> sortof evidence we can use to conclude that the image was a $0$ - wecould legitimately get a $0$ in many other ways (say, throughtranslations of the above images, or slight distortions). But itseems safe to say that at least in this case we'd conclude that theinput was a $0$.</p><p></p><p></p><p></p><p>Supposing the neural network functions in this way, we can give aplausible explanation for why it's better to have $10$ outputs fromthe network, rather than $4$. If we had $4$ outputs, then the firstoutput neuron would be trying to decide what the most significant bitof the digit was. And there's no easy way to relate that mostsignificant bit to simple shapes like those shown above. It's hard toimagine that there's any good historical reason the component shapesof the digit will be closely related to (say) the most significant bitin the output.</p><p>Now, with all that said, this is all just a heuristic. Nothing saysthat the three-layer neural network has to operate in the way Idescribed, with the hidden neurons detecting simple component shapes.Maybe a clever learning algorithm will find some assignment of weightsthat lets us use only $4$ output neurons. But as a heuristic the wayof thinking I've described works pretty well, and can save you a lotof time in designing good neural network architectures.</p><p><h4><a name="exercise_513527"></a><a href="#exercise_513527">Exercise</a></h4><ul><li> There is a way of determining the bitwise representation of a digit by adding an extra layer to the three-layer network above. The extra layer converts the output from the previous layer into a binary representation, as illustrated in the figure below. Find a set of weights and biases for the new output layer. Assume that the first $3$ layers of neurons are such that the correct output in the third layer (i.e., the old output layer) has activation at least $0.99$, and incorrect outputs have activation less than $0.01$.</ul></p><p><center><img src="images/tikz13.png"/></center></p><p></p><p></p><p></p><p><h3><a name="learning_with_gradient_descent"></a><a href="#learning_with_gradient_descent">Learning with gradient descent</a></h3></p><p></p><p>Now that we have a design for our neural network, how can it learn torecognize digits? The first thing we'll need is a data set to learnfrom - a so-called training data set. We'll use the <a href="http://yann.lecun.com/exdb/mnist/">MNIST data set</a>, which contains tens of thousands of scanned images ofhandwritten digits, together with their correct classifications.MNIST's name comes from the fact that it is a modified subset of twodata sets collected by<a href="http://en.wikipedia.org/wiki/National_Institute_of_Standards_and_Technology">NIST</a>,the United States' National Institute of Standards andTechnology. Here's a few images from MNIST:</p><p><center><img src="images/digits_separate.png" width="420px"></center> </p><p>As you can see, these digits are, in fact, the same as those shownat the <a href="#complete_zero">beginning of this chapter</a> as a challengeto recognize. Of course, when testing our network we'll ask it torecognize images which aren't in the training set!</p><p>The MNIST data comes in two parts. The first part contains 60,000images to be used as training data. These images are scannedhandwriting samples from 250 people, half of whom were US CensusBureau employees, and half of whom were high school students. Theimages are greyscale and 28 by 28 pixels in size. The second part ofthe MNIST data set is 10,000 images to be used as test data. Again,these are 28 by 28 greyscale images. We'll use the test data toevaluate how well our neural network has learned to recognize digits.To make this a good test of performance, the test data was taken froma <em>different</em> set of 250 people than the original training data(albeit still a group split between Census Bureau employees and highschool students). This helps give us confidence that our system canrecognize digits from people whose writing it didn't see duringtraining.</p><p>We'll use the notation $x$ to denote a training input. It'll beconvenient to regard each training input $x$ as a $28 \times 28 =784$-dimensional vector. Each entry in the vector represents the greyvalue for a single pixel in the image. We'll denote the correspondingdesired output by $y = y(x)$, where $y$ is a $10$-dimensional vector.For example, if a particular training image, $x$, depicts a $6$, then$y(x) = (0, 0, 0, 0, 0, 0, 1, 0, 0, 0)^T$ is the desired output fromthe network. Note that $T$ here is the transpose operation, turning arow vector into an ordinary (column) vector.</p><p>What we'd like is an algorithm which lets us find weights and biasesso that the output from the network approximates $y(x)$ for alltraining inputs $x$. To quantify how well we're achieving this goalwe define a <em>cost function</em>*<span class="marginnote">*Sometimes referred to as a <em>loss</em> or <em>objective</em> function. We use the term cost function throughout this book, but you should note the other terminology, since it's often used in research papers and other discussions of neural networks. </span>:<a class="displaced_anchor" name="eqtn6"></a>\begin{eqnarray} C(w,b) \equiv \frac{1}{2n} \sum_x \| y(x) - a\|^2.\tag{6}\end{eqnarray}Here, $w$ denotes the collection of all weights in the network, $b$all the biases, $n$ is the total number of training inputs, $a$ is thevector of outputs from the network when $x$ is input, and the sum isover all training inputs, $x$. Of course, the output $a$ depends on$x$, $w$ and $b$, but to keep the notation simple I haven't explicitlyindicated this dependence. The notation $\| v \|$ just denotes theusual length function for a vector $v$. We'll call $C$ the<em>quadratic</em> cost function; it's alsosometimes known as the <em>mean squared error</em> or just <em>MSE</em>.Inspecting the form of the quadratic cost function, we see that$C(w,b)$ is non-negative, since every term in the sum is non-negative.Furthermore, the cost $C(w,b)$ becomes small, i.e., $C(w,b) \approx0$, precisely when $y(x)$ is approximately equal to the output, $a$,for all training inputs, $x$. So our training algorithm has done agood job if it can find weights and biases so that $C(w,b) \approx 0$.By contrast, it's not doing so well when $C(w,b)$ is large - thatwould mean that $y(x)$ is not close to the output $a$ for a largenumber of inputs. So the aim of our training algorithm will be tominimize the cost $C(w,b)$ as a function of the weights and biases.In other words, we want to find a set of weights and biases which makethe cost as small as possible. We'll do that using an algorithm knownas <em>gradient descent</em>.</p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p>Why introduce the quadratic cost? After all, aren't we primarilyinterested in the number of images correctly classified by thenetwork? Why not try to maximize that number directly, rather thanminimizing a proxy measure like the quadratic cost? The problem withthat is that the number of images correctly classified is not a smoothfunction of the weights and biases in the network. For the most part,making small changes to the weights and biases won't cause any changeat all in the number of training images classified correctly. Thatmakes it difficult to figure out how to change the weights and biasesto get improved performance. If we instead use a smooth cost functionlike the quadratic cost it turns out to be easy to figure out how tomake small changes in the weights and biases so as to get animprovement in the cost. That's why we focus first on minimizing thequadratic cost, and only after that will we examine the classificationaccuracy.</p><p></p><p>Even given that we want to use a smooth cost function, you may stillwonder why we choose the quadratic function used inEquation <span id="margin_368924667121_reveal" class="equation_link">(6)</span><span id="margin_368924667121" class="marginequation" style="display: none;"><a href="chap1.html#eqtn6" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} C(w,b) \equiv \frac{1}{2n} \sum_x \| y(x) - a\|^2 \nonumber\end{eqnarray}</a></span><script>$('#margin_368924667121_reveal').click(function() {$('#margin_368924667121').toggle('slow', function() {});});</script>. Isn't this a rather <em>ad hoc</em> choice? Perhaps if we chose a different cost function we'd geta totally different set of minimizing weights and biases? This is avalid concern, and later we'll revisit the cost function, and makesome modifications. However, the quadratic cost function ofEquation <span id="margin_77007455211_reveal" class="equation_link">(6)</span><span id="margin_77007455211" class="marginequation" style="display: none;"><a href="chap1.html#eqtn6" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} C(w,b) \equiv \frac{1}{2n} \sum_x \| y(x) - a\|^2 \nonumber\end{eqnarray}</a></span><script>$('#margin_77007455211_reveal').click(function() {$('#margin_77007455211').toggle('slow', function() {});});</script> works perfectly well forunderstanding the basics of learning in neural networks, so we'llstick with it for now.</p><p>Recapping, our goal in training a neural network is to find weightsand biases which minimize the quadratic cost function $C(w, b)$. Thisis a well-posed problem, but it's got a lot of distracting structureas currently posed - the interpretation of $w$ and $b$ as weightsand biases, the $\sigma$ function lurking in the background, thechoice of network architecture, MNIST, and so on. It turns out thatwe can understand a tremendous amount by ignoring most of thatstructure, and just concentrating on the minimization aspect. So fornow we're going to forget all about the specific form of the costfunction, the connection to neural networks, and so on. Instead,we're going to imagine that we've simply been given a function of manyvariables and we want to minimize that function. We're going todevelop a technique called <em>gradient descent</em> which can be usedto solve such minimization problems. Then we'll come back to thespecific function we want to minimize for neural networks.</p><p>Okay, let's suppose we're trying to minimize some function, $C(v)$.This could be any real-valued function of many variables, $v = v_1,v_2, \ldots$. Note that I've replaced the $w$ and $b$ notation by $v$to emphasize that this could be any function - we're notspecifically thinking in the neural networks context any more. Tominimize $C(v)$ it helps to imagine $C$ as a function of just twovariables, which we'll call $v_1$ and $v_2$:</p><p><center><img src="images/valley.png" width="542px"></center></p><p>What we'd like is to find where $C$ achieves its global minimum. Now,of course, for the function plotted above, we can eyeball the graphand find the minimum. In that sense, I've perhaps shown slightly<em>too</em> simple a function! A general function, $C$, may be acomplicated function of many variables, and it won't usually bepossible to just eyeball the graph to find the minimum.</p><p>One way of attacking the problem is to use calculus to try to find theminimum analytically. We could compute derivatives and then try usingthem to find places where $C$ is an extremum. With some luck thatmight work when $C$ is a function of just one or a few variables. Butit'll turn into a nightmare when we have many more variables. And forneural networks we'll often want <em>far</em> more variables - thebiggest neural networks have cost functions which depend on billionsof weights and biases in an extremely complicated way. Using calculusto minimize that just won't work!</p><p>(After asserting that we'll gain insight by imagining $C$ as afunction of just two variables, I've turned around twice in twoparagraphs and said, "hey, but what if it's a function of many morethan two variables?" Sorry about that. Please believe me when I saythat it really does help to imagine $C$ as a function of twovariables. It just happens that sometimes that picture breaks down,and the last two paragraphs were dealing with such breakdowns. Goodthinking about mathematics often involves juggling multiple intuitivepictures, learning when it's appropriate to use each picture, and whenit's not.)</p><p><a name="gradient_descent"></a></p><p>Okay, so calculus doesn't work. Fortunately, there is a beautifulanalogy which suggests an algorithm which works pretty well. We startby thinking of our function as a kind of a valley. If you squint justa little at the plot above, that shouldn't be too hard. And weimagine a ball rolling down the slope of the valley. Our everydayexperience tells us that the ball will eventually roll to the bottomof the valley. Perhaps we can use this idea as a way to find aminimum for the function? We'd randomly choose a starting point foran (imaginary) ball, and then simulate the motion of the ball as itrolled down to the bottom of the valley. We could do this simulationsimply by computing derivatives (and perhaps some second derivatives)of $C$ - those derivatives would tell us everything we need to knowabout the local "shape" of the valley, and therefore how our ballshould roll.</p><p>Based on what I've just written, you might suppose that we'll betrying to write down Newton's equations of motion for the ball,considering the effects of friction and gravity, and so on. Actually,we're not going to take the ball-rolling analogy quite that seriously- we're devising an algorithm to minimize $C$, not developing anaccurate simulation of the laws of physics! The ball's-eye view ismeant to stimulate our imagination, not constrain our thinking. Sorather than get into all the messy details of physics, let's simplyask ourselves: if we were declared God for a day, and could make upour own laws of physics, dictating to the ball how it should roll,what law or laws of motion could we pick that would make it so theball always rolled to the bottom of the valley?</p><p>To make this question more precise, let's think about what happenswhen we move the ball a small amount $\Delta v_1$ in the $v_1$direction, and a small amount $\Delta v_2$ in the $v_2$ direction.Calculus tells us that $C$ changes as follows:<a class="displaced_anchor" name="eqtn7"></a>\begin{eqnarray} \Delta C \approx \frac{\partial C}{\partial v_1} \Delta v_1 + \frac{\partial C}{\partial v_2} \Delta v_2.\tag{7}\end{eqnarray}We're going to find a way of choosing $\Delta v_1$ and $\Delta v_2$ soas to make $\Delta C$ negative; i.e., we'll choose them so the ball isrolling down into the valley. To figure out how to make such a choiceit helps to define $\Delta v$ to be the vector of changes in $v$,$\Delta v \equiv (\Delta v_1, \Delta v_2)^T$, where $T$ is again thetranspose operation, turning row vectors into column vectors. We'llalso define the <em>gradient</em> of $C$to be the vector of partial derivatives, $\left(\frac{\partial C}{\partial v_1}, \frac{\partial C}{\partial v_2}\right)^T$. Wedenote the gradient vector by $\nabla C$, i.e.:<a class="displaced_anchor" name="eqtn8"></a>\begin{eqnarray} \nabla C \equiv \left( \frac{\partial C}{\partial v_1}, \frac{\partial C}{\partial v_2} \right)^T.\tag{8}\end{eqnarray}In a moment we'll rewrite the change $\Delta C$ in terms of $\Delta v$and the gradient, $\nabla C$. Before getting to that, though, I wantto clarify something that sometimes gets people hung up on thegradient. When meeting the $\nabla C$ notation for the first time,people sometimes wonder how they should think about the $\nabla$symbol. What, exactly, does $\nabla$ mean? In fact, it's perfectlyfine to think of $\nabla C$ as a single mathematical object - thevector defined above - which happens to be written using twosymbols. In this point of view, $\nabla$ is just a piece ofnotational flag-waving, telling you "hey, $\nabla C$ is a gradientvector". There are more advanced points of view where $\nabla$ canbe viewed as an independent mathematical entity in its own right (forexample, as a differential operator), but we won't need such points ofview.</p><p>With these definitions, the expression <span id="margin_60068869945_reveal" class="equation_link">(7)</span><span id="margin_60068869945" class="marginequation" style="display: none;"><a href="chap1.html#eqtn7" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta C \approx \frac{\partial C}{\partial v_1} \Delta v_1 + \frac{\partial C}{\partial v_2} \Delta v_2 \nonumber\end{eqnarray}</a></span><script>$('#margin_60068869945_reveal').click(function() {$('#margin_60068869945').toggle('slow', function() {});});</script> for$\Delta C$ can be rewritten as<a class="displaced_anchor" name="eqtn9"></a>\begin{eqnarray} \Delta C \approx \nabla C \cdot \Delta v.\tag{9}\end{eqnarray}This equation helps explain why $\nabla C$ is called the gradientvector: $\nabla C$ relates changes in $v$ to changes in $C$, just aswe'd expect something called a gradient to do. But what's reallyexciting about the equation is that it lets us see how to choose$\Delta v$ so as to make $\Delta C$ negative. In particular, supposewe choose<a class="displaced_anchor" name="eqtn10"></a>\begin{eqnarray} \Delta v = -\eta \nabla C,\tag{10}\end{eqnarray}where $\eta$ is a small, positive parameter (known as the<em>learning rate</em>).Then Equation <span id="margin_777394057862_reveal" class="equation_link">(9)</span><span id="margin_777394057862" class="marginequation" style="display: none;"><a href="chap1.html#eqtn9" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta C \approx \nabla C \cdot \Delta v \nonumber\end{eqnarray}</a></span><script>$('#margin_777394057862_reveal').click(function() {$('#margin_777394057862').toggle('slow', function() {});});</script> tells us that $\Delta C \approx -\eta\nabla C \cdot \nabla C = -\eta \|\nabla C\|^2$. Because $\| \nabla C\|^2 \geq 0$, this guarantees that $\Delta C \leq 0$, i.e., $C$ willalways decrease, never increase, if we change $v$ according to theprescription in <span id="margin_387482875009_reveal" class="equation_link">(10)</span><span id="margin_387482875009" class="marginequation" style="display: none;"><a href="chap1.html#eqtn10" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta v = -\eta \nabla C \nonumber\end{eqnarray}</a></span><script>$('#margin_387482875009_reveal').click(function() {$('#margin_387482875009').toggle('slow', function() {});});</script>. (Within, of course, thelimits of the approximation in Equation <span id="margin_602571566970_reveal" class="equation_link">(9)</span><span id="margin_602571566970" class="marginequation" style="display: none;"><a href="chap1.html#eqtn9" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta C \approx \nabla C \cdot \Delta v \nonumber\end{eqnarray}</a></span><script>$('#margin_602571566970_reveal').click(function() {$('#margin_602571566970').toggle('slow', function() {});});</script>). This isexactly the property we wanted! And so we'll takeEquation <span id="margin_129183303476_reveal" class="equation_link">(10)</span><span id="margin_129183303476" class="marginequation" style="display: none;"><a href="chap1.html#eqtn10" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta v = -\eta \nabla C \nonumber\end{eqnarray}</a></span><script>$('#margin_129183303476_reveal').click(function() {$('#margin_129183303476').toggle('slow', function() {});});</script> to define the "law of motion"for the ball in our gradient descent algorithm. That is, we'll useEquation <span id="margin_734088671290_reveal" class="equation_link">(10)</span><span id="margin_734088671290" class="marginequation" style="display: none;"><a href="chap1.html#eqtn10" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta v = -\eta \nabla C \nonumber\end{eqnarray}</a></span><script>$('#margin_734088671290_reveal').click(function() {$('#margin_734088671290').toggle('slow', function() {});});</script> to compute a value for $\Deltav$, then move the ball's position $v$ by that amount:<a class="displaced_anchor" name="eqtn11"></a>\begin{eqnarray} v \rightarrow v' = v -\eta \nabla C.\tag{11}\end{eqnarray}Then we'll use this update rule again, to make another move. If wekeep doing this, over and over, we'll keep decreasing $C$ until - wehope - we reach a global minimum.</p><p>Summing up, the way the gradient descent algorithm works is torepeatedly compute the gradient $\nabla C$, and then to move in the<em>opposite</em> direction, "falling down" the slope of the valley.We can visualize it like this:</p><p><center><img src="images/valley_with_ball.png" width="542px"></center></p><p>Notice that with this rule gradient descent doesn't reproduce realphysical motion. In real life a ball has momentum, and that momentummay allow it to roll across the slope, or even (momentarily) rolluphill. It's only after the effects of friction set in that the ballis guaranteed to roll down into the valley. By contrast, our rule forchoosing $\Delta v$ just says "go down, right now". That's still apretty good rule for finding the minimum!</p><p>To make gradient descent work correctly, we need to choose thelearning rate $\eta$ to be smallenough that Equation <span id="margin_693595312216_reveal" class="equation_link">(9)</span><span id="margin_693595312216" class="marginequation" style="display: none;"><a href="chap1.html#eqtn9" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta C \approx \nabla C \cdot \Delta v \nonumber\end{eqnarray}</a></span><script>$('#margin_693595312216_reveal').click(function() {$('#margin_693595312216').toggle('slow', function() {});});</script> is a good approximation. Ifwe don't, we might end up with $\Delta C > 0$, which obviously wouldnot be good! At the same time, we don't want $\eta$ to be too small,since that will make the changes $\Delta v$ tiny, and thus thegradient descent algorithm will work very slowly. In practicalimplementations, $\eta$ is often varied so thatEquation <span id="margin_763885870077_reveal" class="equation_link">(9)</span><span id="margin_763885870077" class="marginequation" style="display: none;"><a href="chap1.html#eqtn9" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta C \approx \nabla C \cdot \Delta v \nonumber\end{eqnarray}</a></span><script>$('#margin_763885870077_reveal').click(function() {$('#margin_763885870077').toggle('slow', function() {});});</script> remains a good approximation, but thealgorithm isn't too slow. We'll see later how thisworks. </p><p>I've explained gradient descent when $C$ is a function of just twovariables. But, in fact, everything works just as well even when $C$is a function of many more variables. Suppose in particular that $C$is a function of $m$ variables, $v_1,\ldots,v_m$. Then the change$\Delta C$ in $C$ produced by a small change $\Delta v = (\Delta v_1,\ldots, \Delta v_m)^T$ is<a class="displaced_anchor" name="eqtn12"></a>\begin{eqnarray} \Delta C \approx \nabla C \cdot \Delta v,\tag{12}\end{eqnarray}where the gradient $\nabla C$ is the vector <a class="displaced_anchor" name="eqtn13"></a>\begin{eqnarray} \nabla C \equiv \left(\frac{\partial C}{\partial v_1}, \ldots, \frac{\partial C}{\partial v_m}\right)^T.\tag{13}\end{eqnarray}Just as for the two variable case, we canchoose<a class="displaced_anchor" name="eqtn14"></a>\begin{eqnarray} \Delta v = -\eta \nabla C,\tag{14}\end{eqnarray}and we're guaranteed that our (approximate)expression <span id="margin_796021234053_reveal" class="equation_link">(12)</span><span id="margin_796021234053" class="marginequation" style="display: none;"><a href="chap1.html#eqtn12" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \Delta C \approx \nabla C \cdot \Delta v \nonumber\end{eqnarray}</a></span><script>$('#margin_796021234053_reveal').click(function() {$('#margin_796021234053').toggle('slow', function() {});});</script> for $\Delta C$ will be negative.This gives us a way of following the gradient to a minimum, even when$C$ is a function of many variables, by repeatedly applying the updaterule<a class="displaced_anchor" name="eqtn15"></a>\begin{eqnarray} v \rightarrow v' = v-\eta \nabla C.\tag{15}\end{eqnarray}You can think of this update rule as <em>defining</em> the gradientdescent algorithm. It gives us a way of repeatedly changing theposition $v$ in order to find a minimum of the function $C$. The ruledoesn't always work - several things can go wrong and preventgradient descent from finding the global minimum of $C$, a point we'llreturn to explore in later chapters. But, in practice gradientdescent often works extremely well, and in neural networks we'll findthat it's a powerful way of minimizing the cost function, and sohelping the net learn.</p><p></p><p></p><p>Indeed, there's even a sense in which gradient descent is the optimalstrategy for searching for a minimum. Let's suppose that we're tryingto make a move $\Delta v$ in position so as to decrease $C$ as much aspossible. This is equivalent to minimizing $\Delta C \approx \nabla C\cdot \Delta v$. We'll constrain the size of the move so that $\|\Delta v \| = \epsilon$ for some small fixed $\epsilon > 0$. In otherwords, we want a move that is a small step of a fixed size, and we'retrying to find the movement direction which decreases $C$ as much aspossible. It can be proved that the choice of $\Delta v$ whichminimizes $\nabla C \cdot \Delta v$ is $\Delta v = - \eta \nabla C$,where $\eta = \epsilon / \|\nabla C\|$ is determined by the sizeconstraint $\|\Delta v\| = \epsilon$. So gradient descent can beviewed as a way of taking small steps in the direction which does themost to immediately decrease $C$.</p><p><h4><a name="exercises_647181"></a><a href="#exercises_647181">Exercises</a></h4><ul><li> Prove the assertion of the last paragraph. <em>Hint:</em> If you're not already familiar with the <a href="http://en.wikipedia.org/wiki/Cauchy%E2%80%93Schwarz_inequality">Cauchy-Schwarz inequality</a>, you may find it helpful to familiarize yourself with it.</p><p><li> I explained gradient descent when $C$ is a function of two variables, and when it's a function of more than two variables. What happens when $C$ is a function of just one variable? Can you provide a geometric interpretation of what gradient descent is doing in the one-dimensional case?</ul></p><p></p><p>People have investigated many variations of gradient descent,including variations that more closely mimic a real physical ball.These ball-mimicking variations have some advantages, but also have amajor disadvantage: it turns out to be necessary to compute secondpartial derivatives of $C$, and this can be quite costly. To see whyit's costly, suppose we want to compute all the second partialderivatives $\partial^2 C/ \partial v_j \partial v_k$. If there are amillion such $v_j$ variables then we'd need to compute something likea trillion (i.e., a million squared) second partialderivatives*<span class="marginnote">*Actually, more like half a trillion, since $\partial^2 C/ \partial v_j \partial v_k = \partial^2 C/ \partial v_k \partial v_j$. Still, you get the point.</span>! That's going to becomputationally costly. With that said, there are tricks for avoidingthis kind of problem, and finding alternatives to gradient descent isan active area of investigation. But in this book we'll use gradientdescent (and variations) as our main approach to learning in neuralnetworks.</p><p>How can we apply gradient descent to learn in a neural network? Theidea is to use gradient descent to find the weights $w_k$ and biases$b_l$ which minimize the cost inEquation <span id="margin_552678515184_reveal" class="equation_link">(6)</span><span id="margin_552678515184" class="marginequation" style="display: none;"><a href="chap1.html#eqtn6" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} C(w,b) \equiv \frac{1}{2n} \sum_x \| y(x) - a\|^2 \nonumber\end{eqnarray}</a></span><script>$('#margin_552678515184_reveal').click(function() {$('#margin_552678515184').toggle('slow', function() {});});</script>. To see how this works, let'srestate the gradient descent update rule, with the weights and biasesreplacing the variables $v_j$. In other words, our "position" nowhas components $w_k$ and $b_l$, and the gradient vector $\nabla C$ hascorresponding components $\partial C / \partial w_k$ and $\partial C/ \partial b_l$. Writing out the gradient descent update rule interms of components, we have<a class="displaced_anchor" name="eqtn16"></a><a class="displaced_anchor" name="eqtn17"></a>\begin{eqnarray} w_k & \rightarrow & w_k' = w_k-\eta \frac{\partial C}{\partial w_k} \tag{16}\\ b_l & \rightarrow & b_l' = b_l-\eta \frac{\partial C}{\partial b_l}.\tag{17}\end{eqnarray}By repeatedly applying this update rule we can "roll down the hill",and hopefully find a minimum of the cost function. In other words,this is a rule which can be used to learn in a neural network.</p><p>There are a number of challenges in applying the gradient descentrule. We'll look into those in depth in later chapters. But for nowI just want to mention one problem. To understand what the problemis, let's look back at the quadratic cost inEquation <span id="margin_636312544623_reveal" class="equation_link">(6)</span><span id="margin_636312544623" class="marginequation" style="display: none;"><a href="chap1.html#eqtn6" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} C(w,b) \equiv \frac{1}{2n} \sum_x \| y(x) - a\|^2 \nonumber\end{eqnarray}</a></span><script>$('#margin_636312544623_reveal').click(function() {$('#margin_636312544623').toggle('slow', function() {});});</script>. Notice that this costfunction has the form $C = \frac{1}{n} \sum_x C_x$, that is, it's anaverage over costs $C_x \equiv \frac{\|y(x)-a\|^2}{2}$ for individualtraining examples. In practice, to compute the gradient $\nabla C$ weneed to compute the gradients $\nabla C_x$ separately for eachtraining input, $x$, and then average them, $\nabla C = \frac{1}{n}\sum_x \nabla C_x$. Unfortunately, when the number of training inputsis very large this can take a long time, and learning thus occursslowly.</p><p>An idea called <em>stochastic gradient descent</em> can be used to speedup learning. The idea is to estimate the gradient $\nabla C$ bycomputing $\nabla C_x$ for a small sample of randomly chosen traininginputs. By averaging over this small sample it turns out that we canquickly get a good estimate of the true gradient $\nabla C$, and thishelps speed up gradient descent, and thus learning.</p><p>To make these ideas more precise, stochastic gradient descent works byrandomly picking out a small number $m$ of randomly chosen traininginputs. We'll label those random training inputs $X_1, X_2, \ldots,X_m$, and refer to them as a <em>mini-batch</em>. Provided the samplesize $m$ is large enough we expect that the average value of the$\nabla C_{X_j}$ will be roughly equal to the average over all $\nablaC_x$, that is,<a class="displaced_anchor" name="eqtn18"></a>\begin{eqnarray} \frac{\sum_{j=1}^m \nabla C_{X_{j}}}{m} \approx \frac{\sum_x \nabla C_x}{n} = \nabla C,\tag{18}\end{eqnarray}where the second sum is over the entire set of training data.Swapping sides we get<a class="displaced_anchor" name="eqtn19"></a>\begin{eqnarray} \nabla C \approx \frac{1}{m} \sum_{j=1}^m \nabla C_{X_{j}},\tag{19}\end{eqnarray}confirming that we can estimate the overall gradient by computinggradients just for the randomly chosen mini-batch. </p><p>To connect this explicitly to learning in neural networks, suppose$w_k$ and $b_l$ denote the weights and biases in our neural network.Then stochastic gradient descent works by picking out a randomlychosen mini-batch of training inputs, and training with those,<a class="displaced_anchor" name="eqtn20"></a><a class="displaced_anchor" name="eqtn21"></a>\begin{eqnarray} w_k & \rightarrow & w_k' = w_k-\frac{\eta}{m} \sum_j \frac{\partial C_{X_j}}{\partial w_k} \tag{20}\\ b_l & \rightarrow & b_l' = b_l-\frac{\eta}{m} \sum_j \frac{\partial C_{X_j}}{\partial b_l},\tag{21}\end{eqnarray}where the sums are over all the training examples $X_j$ in the currentmini-batch. Then we pick out another randomly chosen mini-batch andtrain with those. And so on, until we've exhausted the traininginputs, which is said to complete an<em>epoch</em> of training. At that pointwe start over with a new training epoch.</p><p>Incidentally, it's worth noting that conventions vary about scaling ofthe cost function and of mini-batch updates to the weights and biases.In Equation <span id="margin_28961100271_reveal" class="equation_link">(6)</span><span id="margin_28961100271" class="marginequation" style="display: none;"><a href="chap1.html#eqtn6" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} C(w,b) \equiv \frac{1}{2n} \sum_x \| y(x) - a\|^2 \nonumber\end{eqnarray}</a></span><script>$('#margin_28961100271_reveal').click(function() {$('#margin_28961100271').toggle('slow', function() {});});</script> we scaled the overall costfunction by a factor $\frac{1}{n}$. People sometimes omit the$\frac{1}{n}$, summing over the costs of individual training examplesinstead of averaging. This is particularly useful when the totalnumber of training examples isn't known in advance. This can occur ifmore training data is being generated in real time, for instance.And, in a similar way, the mini-batch update rules <span id="margin_38667351831_reveal" class="equation_link">(20)</span><span id="margin_38667351831" class="marginequation" style="display: none;"><a href="chap1.html#eqtn20" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} w_k & \rightarrow & w_k' = w_k-\frac{\eta}{m} \sum_j \frac{\partial C_{X_j}}{\partial w_k} \nonumber\end{eqnarray}</a></span><script>$('#margin_38667351831_reveal').click(function() {$('#margin_38667351831').toggle('slow', function() {});});</script>and <span id="margin_667554963539_reveal" class="equation_link">(21)</span><span id="margin_667554963539" class="marginequation" style="display: none;"><a href="chap1.html#eqtn21" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} b_l & \rightarrow & b_l' = b_l-\frac{\eta}{m} \sum_j \frac{\partial C_{X_j}}{\partial b_l} \nonumber\end{eqnarray}</a></span><script>$('#margin_667554963539_reveal').click(function() {$('#margin_667554963539').toggle('slow', function() {});});</script> sometimes omit the $\frac{1}{m}$ term out thefront of the sums. Conceptually this makes little difference, sinceit's equivalent to rescaling the learning rate $\eta$. But when doingdetailed comparisons of different work it's worth watching out for.</p><p>We can think of stochastic gradient descent as being like politicalpolling: it's much easier to sample a small mini-batch than it is toapply gradient descent to the full batch, just as carrying out a pollis easier than running a full election. For example, if we have atraining set of size $n = 60,000$, as in MNIST, and choose amini-batch size of (say) $m = 10$, this means we'll get a factor of$6,000$ speedup in estimating the gradient! Of course, the estimatewon't be perfect - there will be statistical fluctuations - but itdoesn't need to be perfect: all we really care about is moving in ageneral direction that will help decrease $C$, and that means we don'tneed an exact computation of the gradient. In practice, stochasticgradient descent is a commonly used and powerful technique forlearning in neural networks, and it's the basis for most of thelearning techniques we'll develop in this book.</p><p></p><p></p><p></p><p></p><p></p><p><h4><a name="exercise_263792"></a><a href="#exercise_263792">Exercise</a></h4><ul><li> An extreme version of gradient descent is to use a mini-batch size of just 1. That is, given a training input, $x$, we update our weights and biases according to the rules $w_k \rightarrow w_k' = w_k - \eta \partial C_x / \partial w_k$ and $b_l \rightarrow b_l' = b_l - \eta \partial C_x / \partial b_l$. Then we choose another training input, and update the weights and biases again. And so on, repeatedly. This procedure is known as <em>online</em>, <em>on-line</em>, or <em>incremental</em> learning. In online learning, a neural network learns from just one training input at a time (just as human beings do). Name one advantage and one disadvantage of online learning, compared to stochastic gradient descent with a mini-batch size of, say, $20$.</ul></p><p>Let me conclude this section by discussing a point that sometimes bugspeople new to gradient descent. In neural networks the cost $C$ is,of course, a function of many variables - all the weights and biases- and so in some sense defines a surface in a very high-dimensionalspace. Some people get hung up thinking: "Hey, I have to be able tovisualize all these extra dimensions". And they may start to worry:"I can't think in four dimensions, let alone five (or fivemillion)". Is there some special ability they're missing, someability that "real" supermathematicians have? Of course, the answeris no. Even most professional mathematicians can't visualize fourdimensions especially well, if at all. The trick they use, instead,is to develop other ways of representing what's going on. That'sexactly what we did above: we used an algebraic (rather than visual)representation of $\Delta C$ to figure out how to move so as todecrease $C$. People who are good at thinking in high dimensions havea mental library containing many different techniques along theselines; our algebraic trick is just one example. Those techniques maynot have the simplicity we're accustomed to when visualizing threedimensions, but once you build up a library of such techniques, youcan get pretty good at thinking in high dimensions. I won't go intomore detail here, but if you're interested then you may enjoy reading<a href="http://mathoverflow.net/questions/25983/intuitive-crutches-for-higher-dimensional-thinking">this discussion</a> of some of the techniques professional mathematiciansuse to think in high dimensions. While some of the techniquesdiscussed are quite complex, much of the best content is intuitive andaccessible, and could be mastered by anyone.</p><p></p><p><h3><a name="implementing_our_network_to_classify_digits"></a><a href="#implementing_our_network_to_classify_digits">Implementing our network to classify digits</a></h3></p><p>Alright, let's write a program that learns how to recognizehandwritten digits, using stochastic gradient descent and the MNISTtraining data. We'll do this with a short Python (2.7) program, just74 lines of code! The first thing we need is to get the MNIST data.If you're a <tt>git</tt> user then you can obtain the data by cloningthe code repository for this book,</p><p><div class="highlight"><pre><span></span>git clone https://github.com/mnielsen/neural-networks-and-deep-learning.git
</pre></div>
</p><p>If you don't use <tt>git</tt> then you can download the data and code<a href="https://github.com/mnielsen/neural-networks-and-deep-learning/archive/master.zip">here</a>.</p><p>Incidentally, when I described the MNIST data earlier, I said it wassplit into 60,000 training images, and 10,000 test images. That's theofficial MNIST description. Actually, we're going to split the data alittle differently. We'll leave the test images as is, but split the60,000-image MNIST training set into two parts: a set of 50,000images, which we'll use to train our neural network, and a separate10,000 image <em>validation set</em>. We won'tuse the validation data in this chapter, but later in the book we'llfind it useful in figuring out how to set certain<em>hyper-parameters</em> of the neural network - things like thelearning rate, and so on, which aren't directly selected by ourlearning algorithm. Although the validation data isn't part of theoriginal MNIST specification, many people use MNIST in this fashion,and the use of validation data is common in neural networks. When Irefer to the "MNIST training data" from now on, I'll be referring toour 50,000 image data set, not the original 60,000 image dataset*<span class="marginnote">*As noted earlier, the MNIST data set is based on two data sets collected by NIST, the United States' National Institute of Standards and Technology. To construct MNIST the NIST data sets were stripped down and put into a more convenient format by Yann LeCun, Corinna Cortes, and Christopher J. C. Burges. See <a href="http://yann.lecun.com/exdb/mnist/">this link</a> for more details. The data set in my repository is in a form that makes it easy to load and manipulate the MNIST data in Python. I obtained this particular form of the data from the LISA machine learning laboratory at the University of Montreal (<a href="http://www.deeplearning.net/tutorial/gettingstarted.html">link</a>).</span>.</p><p></p><p>Apart from the MNIST data we also need a Python library called<a href="http://numpy.org">Numpy</a>, for doing fast linear algebra. If youdon't already have Numpy installed, you can get it<a href="http://www.scipy.org/install.html">here</a>.</p><p>Let me explain the core features of the neural networks code, beforegiving a full listing, below. The centerpiece is a <tt>Network</tt>class, which we use to represent a neural network. Here's the code weuse to initialize a <tt>Network</tt> object:</p><p><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">Network</span><span class="p">(</span><span class="nb">object</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">sizes</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="n">sizes</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">sizes</span> <span class="o">=</span> <span class="n">sizes</span>
<span class="bp">self</span><span class="o">.</span><span class="n">biases</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">random</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="n">y</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span> <span class="k">for</span> <span class="n">y</span> <span class="ow">in</span> <span class="n">sizes</span><span class="p">[</span><span class="mi">1</span><span class="p">:]]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">weights</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">random</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="n">y</span><span class="p">,</span> <span class="n">x</span><span class="p">)</span>
<span class="k">for</span> <span class="n">x</span><span class="p">,</span> <span class="n">y</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">sizes</span><span class="p">[:</span><span class="o">-</span><span class="mi">1</span><span class="p">],</span> <span class="n">sizes</span><span class="p">[</span><span class="mi">1</span><span class="p">:])]</span>
</pre></div>
</p><p>In this code, the list <tt>sizes</tt> contains the number of neurons inthe respective layers. So, for example, if we want to create a<tt>Network</tt> object with 2 neurons in the first layer, 3 neurons inthe second layer, and 1 neuron in the final layer, we'd do this withthe code:<div class="highlight"><pre><span></span><span class="n">net</span> <span class="o">=</span> <span class="n">Network</span><span class="p">([</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
</pre></div>
<a name="weight_initialization"></a> The biasesand weights in the <tt>Network</tt> object are all initialized randomly,using the Numpy <tt>np.random.randn</tt> function to generate Gaussiandistributions with mean $0$ and standard deviation $1$. This randominitialization gives our stochastic gradient descent algorithm a placeto start from. In later chapters we'll find better ways ofinitializing the weights and biases, but this will do for now. Notethat the <tt>Network</tt> initialization code assumes that the firstlayer of neurons is an input layer, and omits to set any biases forthose neurons, since biases are only ever used in computing theoutputs from later layers.</p><p>Note also that the biases and weights are stored as lists of Numpymatrices. So, for example <tt>net.weights[1]</tt> is a Numpy matrixstoring the weights connecting the second and third layers of neurons.(It's not the first and second layers, since Python's list indexingstarts at <tt>0</tt>.) Since <tt>net.weights[1]</tt> is rather verbose,let's just denote that matrix $w$. It's a matrix such that $w_{jk}$is the weight for the connection between the $k^{\rm th}$ neuron in thesecond layer, and the $j^{\rm th}$ neuron in the third layer. This orderingof the $j$ and $k$ indices may seem strange - surely it'd make moresense to swap the $j$ and $k$ indices around? The big advantage ofusing this ordering is that it means that the vector of activations ofthe third layer of neurons is:<a class="displaced_anchor" name="eqtn22"></a>\begin{eqnarray} a' = \sigma(w a + b).\tag{22}\end{eqnarray}There's quite a bit going on in this equation, so let's unpack itpiece by piece. $a$ is the vector of activations of the second layerof neurons. To obtain $a'$ we multiply $a$ by the weight matrix $w$,and add the vector $b$ of biases. We then apply the function $\sigma$elementwise to every entry in the vector $w a +b$. (This is called<em>vectorizing</em> the function$\sigma$.) It's easy to verify thatEquation <span id="margin_469346701810_reveal" class="equation_link">(22)</span><span id="margin_469346701810" class="marginequation" style="display: none;"><a href="chap1.html#eqtn22" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} a' = \sigma(w a + b) \nonumber\end{eqnarray}</a></span><script>$('#margin_469346701810_reveal').click(function() {$('#margin_469346701810').toggle('slow', function() {});});</script> gives the same result as ourearlier rule, Equation <span id="margin_803037168757_reveal" class="equation_link">(4)</span><span id="margin_803037168757" class="marginequation" style="display: none;"><a href="chap1.html#eqtn4" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \frac{1}{1+\exp(-\sum_j w_j x_j-b)} \nonumber\end{eqnarray}</a></span><script>$('#margin_803037168757_reveal').click(function() {$('#margin_803037168757').toggle('slow', function() {});});</script>, forcomputing the output of a sigmoid neuron.</p><p><h4><a name="exercise_717502"></a><a href="#exercise_717502">Exercise</a></h4><ul><li> Write out Equation <span id="margin_585248828587_reveal" class="equation_link">(22)</span><span id="margin_585248828587" class="marginequation" style="display: none;"><a href="chap1.html#eqtn22" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} a' = \sigma(w a + b) \nonumber\end{eqnarray}</a></span><script>$('#margin_585248828587_reveal').click(function() {$('#margin_585248828587').toggle('slow', function() {});});</script> in component form, and verify that it gives the same result as the rule <span id="margin_208193369319_reveal" class="equation_link">(4)</span><span id="margin_208193369319" class="marginequation" style="display: none;"><a href="chap1.html#eqtn4" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} \frac{1}{1+\exp(-\sum_j w_j x_j-b)} \nonumber\end{eqnarray}</a></span><script>$('#margin_208193369319_reveal').click(function() {$('#margin_208193369319').toggle('slow', function() {});});</script> for computing the output of a sigmoid neuron.</ul></p><p>With all this in mind, it's easy to write code computing the outputfrom a <tt>Network</tt> instance. We begin by defining the sigmoidfunction:<div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">sigmoid</span><span class="p">(</span><span class="n">z</span><span class="p">):</span>
<span class="k">return</span> <span class="mf">1.0</span><span class="o">/</span><span class="p">(</span><span class="mf">1.0</span><span class="o">+</span><span class="n">np</span><span class="o">.</span><span class="n">exp</span><span class="p">(</span><span class="o">-</span><span class="n">z</span><span class="p">))</span>
</pre></div>
Note that when the input <tt>z</tt> is a vector or Numpy array, Numpyautomatically applies the function <tt>sigmoid</tt> elementwise, thatis, in vectorized form.</p><p>We then add a <tt>feedforward</tt> method to the <tt>Network</tt> class,which, given an input <tt>a</tt> for the network, returns thecorresponding output*<span class="marginnote">*It is assumed that the input <tt>a</tt> is an <tt>(n, 1)</tt> Numpy ndarray, not a <tt>(n,)</tt> vector. Here, <tt>n</tt> is the number of inputs to the network. If you try to use an <tt>(n,)</tt> vector as input you'll get strange results. Although using an <tt>(n,)</tt> vector appears the more natural choice, using an <tt>(n, 1)</tt> ndarray makes it particularly easy to modify the code to feedforward multiple inputs at once, and that is sometimes convenient. </span>. All the method does is appliesEquation <span id="margin_436898280460_reveal" class="equation_link">(22)</span><span id="margin_436898280460" class="marginequation" style="display: none;"><a href="chap1.html#eqtn22" style="padding-bottom: 5px;" onMouseOver="this.style.borderBottom='1px solid #2A6EA6';" onMouseOut="this.style.borderBottom='0px';">\begin{eqnarray} a' = \sigma(w a + b) \nonumber\end{eqnarray}</a></span><script>$('#margin_436898280460_reveal').click(function() {$('#margin_436898280460').toggle('slow', function() {});});</script> for each layer:<div class="highlight"><pre><span></span> <span class="k">def</span> <span class="nf">feedforward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">a</span><span class="p">):</span>
<span class="sd">"""Return the output of the network if "a" is input."""</span>
<span class="k">for</span> <span class="n">b</span><span class="p">,</span> <span class="n">w</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">):</span>
<span class="n">a</span> <span class="o">=</span> <span class="n">sigmoid</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">a</span><span class="p">)</span><span class="o">+</span><span class="n">b</span><span class="p">)</span>
<span class="k">return</span> <span class="n">a</span>
</pre></div>
</p><p>Of course, the main thing we want our <tt>Network</tt> objects to do isto learn. To that end we'll give them an <tt>SGD</tt> method whichimplements stochastic gradient descent. Here's the code. It's alittle mysterious in a few places, but I'll break it down below, afterthe listing.</p><p><div class="highlight"><pre><span></span> <span class="k">def</span> <span class="nf">SGD</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">training_data</span><span class="p">,</span> <span class="n">epochs</span><span class="p">,</span> <span class="n">mini_batch_size</span><span class="p">,</span> <span class="n">eta</span><span class="p">,</span>
<span class="n">test_data</span><span class="o">=</span><span class="bp">None</span><span class="p">):</span>
<span class="sd">"""Train the neural network using mini-batch stochastic</span>
<span class="sd"> gradient descent. The "training_data" is a list of tuples</span>
<span class="sd"> "(x, y)" representing the training inputs and the desired</span>
<span class="sd"> outputs. The other non-optional parameters are</span>
<span class="sd"> self-explanatory. If "test_data" is provided then the</span>
<span class="sd"> network will be evaluated against the test data after each</span>
<span class="sd"> epoch, and partial progress printed out. This is useful for</span>
<span class="sd"> tracking progress, but slows things down substantially."""</span>
<span class="k">if</span> <span class="n">test_data</span><span class="p">:</span> <span class="n">n_test</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="n">test_data</span><span class="p">)</span>
<span class="n">n</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="n">training_data</span><span class="p">)</span>
<span class="k">for</span> <span class="n">j</span> <span class="ow">in</span> <span class="nb">xrange</span><span class="p">(</span><span class="n">epochs</span><span class="p">):</span>
<span class="n">random</span><span class="o">.</span><span class="n">shuffle</span><span class="p">(</span><span class="n">training_data</span><span class="p">)</span>
<span class="n">mini_batches</span> <span class="o">=</span> <span class="p">[</span>
<span class="n">training_data</span><span class="p">[</span><span class="n">k</span><span class="p">:</span><span class="n">k</span><span class="o">+</span><span class="n">mini_batch_size</span><span class="p">]</span>
<span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="nb">xrange</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">n</span><span class="p">,</span> <span class="n">mini_batch_size</span><span class="p">)]</span>
<span class="k">for</span> <span class="n">mini_batch</span> <span class="ow">in</span> <span class="n">mini_batches</span><span class="p">:</span>
<span class="bp">self</span><span class="o">.</span><span class="n">update_mini_batch</span><span class="p">(</span><span class="n">mini_batch</span><span class="p">,</span> <span class="n">eta</span><span class="p">)</span>
<span class="k">if</span> <span class="n">test_data</span><span class="p">:</span>
<span class="k">print</span> <span class="s2">"Epoch {0}: {1} / {2}"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span>
<span class="n">j</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">evaluate</span><span class="p">(</span><span class="n">test_data</span><span class="p">),</span> <span class="n">n_test</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">print</span> <span class="s2">"Epoch {0} complete"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">j</span><span class="p">)</span>
</pre></div>
</p><p>The <tt>training_data</tt> is a list of tuples <tt>(x, y)</tt>representing the training inputs and corresponding desired outputs.The variables <tt>epochs</tt> and <tt>mini_batch_size</tt> are what you'dexpect - the number of epochs to train for, and the size of themini-batches to use when sampling. <tt>eta</tt> is the learning rate,$\eta$. If the optional argument <tt>test_data</tt> is supplied, thenthe program will evaluate the network after each epoch of training,and print out partial progress. This is useful for tracking progress,but slows things down substantially.</p><p>The code works as follows. In each epoch, it starts by randomlyshuffling the training data, and then partitions it into mini-batchesof the appropriate size. This is an easy way of sampling randomlyfrom the training data. Then for each <tt>mini_batch</tt> we apply asingle step of gradient descent. This is done by the code<tt>self.update_mini_batch(mini_batch, eta)</tt>, which updates thenetwork weights and biases according to a single iteration of gradientdescent, using just the training data in <tt>mini_batch</tt>. Here'sthe code for the <tt>update_mini_batch</tt> method:<div class="highlight"><pre><span></span> <span class="k">def</span> <span class="nf">update_mini_batch</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">mini_batch</span><span class="p">,</span> <span class="n">eta</span><span class="p">):</span>
<span class="sd">"""Update the network's weights and biases by applying</span>
<span class="sd"> gradient descent using backpropagation to a single mini batch.</span>
<span class="sd"> The "mini_batch" is a list of tuples "(x, y)", and "eta"</span>
<span class="sd"> is the learning rate."""</span>
<span class="n">nabla_b</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">b</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="k">for</span> <span class="n">b</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">]</span>
<span class="n">nabla_w</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">w</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="k">for</span> <span class="n">w</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">]</span>
<span class="k">for</span> <span class="n">x</span><span class="p">,</span> <span class="n">y</span> <span class="ow">in</span> <span class="n">mini_batch</span><span class="p">:</span>
<span class="n">delta_nabla_b</span><span class="p">,</span> <span class="n">delta_nabla_w</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">backprop</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span>
<span class="n">nabla_b</span> <span class="o">=</span> <span class="p">[</span><span class="n">nb</span><span class="o">+</span><span class="n">dnb</span> <span class="k">for</span> <span class="n">nb</span><span class="p">,</span> <span class="n">dnb</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">nabla_b</span><span class="p">,</span> <span class="n">delta_nabla_b</span><span class="p">)]</span>
<span class="n">nabla_w</span> <span class="o">=</span> <span class="p">[</span><span class="n">nw</span><span class="o">+</span><span class="n">dnw</span> <span class="k">for</span> <span class="n">nw</span><span class="p">,</span> <span class="n">dnw</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">nabla_w</span><span class="p">,</span> <span class="n">delta_nabla_w</span><span class="p">)]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">weights</span> <span class="o">=</span> <span class="p">[</span><span class="n">w</span><span class="o">-</span><span class="p">(</span><span class="n">eta</span><span class="o">/</span><span class="nb">len</span><span class="p">(</span><span class="n">mini_batch</span><span class="p">))</span><span class="o">*</span><span class="n">nw</span>
<span class="k">for</span> <span class="n">w</span><span class="p">,</span> <span class="n">nw</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">,</span> <span class="n">nabla_w</span><span class="p">)]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">biases</span> <span class="o">=</span> <span class="p">[</span><span class="n">b</span><span class="o">-</span><span class="p">(</span><span class="n">eta</span><span class="o">/</span><span class="nb">len</span><span class="p">(</span><span class="n">mini_batch</span><span class="p">))</span><span class="o">*</span><span class="n">nb</span>
<span class="k">for</span> <span class="n">b</span><span class="p">,</span> <span class="n">nb</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">,</span> <span class="n">nabla_b</span><span class="p">)]</span>
</pre></div>
Most of the work is done by the line<div class="highlight"><pre><span></span> <span class="n">delta_nabla_b</span><span class="p">,</span> <span class="n">delta_nabla_w</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">backprop</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span>
</pre></div>
This invokes something called the <em>backpropagation</em> algorithm,which is a fast way of computing the gradient of the cost function.So <tt>update_mini_batch</tt> works simply by computing these gradientsfor every training example in the <tt>mini_batch</tt>, and then updating<tt>self.weights</tt> and <tt>self.biases</tt> appropriately.</p><p>I'm not going to show the code for <tt>self.backprop</tt> right now.We'll study how backpropagation works in the next chapter, includingthe code for <tt>self.backprop</tt>. For now, just assume that itbehaves as claimed, returning the appropriate gradient for the costassociated to the training example <tt>x</tt>.</p><p>Let's look at the full program, including the documentation strings,which I omitted above. Apart from <tt>self.backprop</tt> the program isself-explanatory - all the heavy lifting is done in <tt>self.SGD</tt>and <tt>self.update_mini_batch</tt>, which we've already discussed. The<tt>self.backprop</tt> method makes use of a few extra functions to helpin computing the gradient, namely <tt>sigmoid_prime</tt>, which computesthe derivative of the $\sigma$ function, and<tt>self.cost_derivative</tt>, which I won't describe here. You can getthe gist of these (and perhaps the details) just by looking at thecode and documentation strings. We'll look at them in detail in thenext chapter. Note that while the program appears lengthy, much of the code isdocumentation strings intended to make the code easy to understand.In fact, the program contains just 74 lines of non-whitespace,non-comment code. All the code may be found on GitHub<a href="https://github.com/mnielsen/neural-networks-and-deep-learning/blob/master/src/network.py">here</a>.</p><p></p><p><div class="highlight"><pre><span></span><span class="sd">"""</span>
<span class="sd">network.py</span>
<span class="sd">~~~~~~~~~~</span>
<span class="sd">A module to implement the stochastic gradient descent learning</span>
<span class="sd">algorithm for a feedforward neural network. Gradients are calculated</span>
<span class="sd">using backpropagation. Note that I have focused on making the code</span>
<span class="sd">simple, easily readable, and easily modifiable. It is not optimized,</span>
<span class="sd">and omits many desirable features.</span>
<span class="sd">"""</span>
<span class="c1">#### Libraries</span>
<span class="c1"># Standard library</span>
<span class="kn">import</span> <span class="nn">random</span>
<span class="c1"># Third-party libraries</span>
<span class="kn">import</span> <span class="nn">numpy</span> <span class="kn">as</span> <span class="nn">np</span>
<span class="k">class</span> <span class="nc">Network</span><span class="p">(</span><span class="nb">object</span><span class="p">):</span>
<span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">sizes</span><span class="p">):</span>
<span class="sd">"""The list ``sizes`` contains the number of neurons in the</span>
<span class="sd"> respective layers of the network. For example, if the list</span>
<span class="sd"> was [2, 3, 1] then it would be a three-layer network, with the</span>
<span class="sd"> first layer containing 2 neurons, the second layer 3 neurons,</span>
<span class="sd"> and the third layer 1 neuron. The biases and weights for the</span>
<span class="sd"> network are initialized randomly, using a Gaussian</span>
<span class="sd"> distribution with mean 0, and variance 1. Note that the first</span>
<span class="sd"> layer is assumed to be an input layer, and by convention we</span>
<span class="sd"> won't set any biases for those neurons, since biases are only</span>
<span class="sd"> ever used in computing the outputs from later layers."""</span>
<span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="n">sizes</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">sizes</span> <span class="o">=</span> <span class="n">sizes</span>
<span class="bp">self</span><span class="o">.</span><span class="n">biases</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">random</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="n">y</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span> <span class="k">for</span> <span class="n">y</span> <span class="ow">in</span> <span class="n">sizes</span><span class="p">[</span><span class="mi">1</span><span class="p">:]]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">weights</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">random</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="n">y</span><span class="p">,</span> <span class="n">x</span><span class="p">)</span>
<span class="k">for</span> <span class="n">x</span><span class="p">,</span> <span class="n">y</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">sizes</span><span class="p">[:</span><span class="o">-</span><span class="mi">1</span><span class="p">],</span> <span class="n">sizes</span><span class="p">[</span><span class="mi">1</span><span class="p">:])]</span>
<span class="k">def</span> <span class="nf">feedforward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">a</span><span class="p">):</span>
<span class="sd">"""Return the output of the network if ``a`` is input."""</span>
<span class="k">for</span> <span class="n">b</span><span class="p">,</span> <span class="n">w</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">):</span>
<span class="n">a</span> <span class="o">=</span> <span class="n">sigmoid</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">a</span><span class="p">)</span><span class="o">+</span><span class="n">b</span><span class="p">)</span>
<span class="k">return</span> <span class="n">a</span>
<span class="k">def</span> <span class="nf">SGD</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">training_data</span><span class="p">,</span> <span class="n">epochs</span><span class="p">,</span> <span class="n">mini_batch_size</span><span class="p">,</span> <span class="n">eta</span><span class="p">,</span>
<span class="n">test_data</span><span class="o">=</span><span class="bp">None</span><span class="p">):</span>
<span class="sd">"""Train the neural network using mini-batch stochastic</span>
<span class="sd"> gradient descent. The ``training_data`` is a list of tuples</span>
<span class="sd"> ``(x, y)`` representing the training inputs and the desired</span>
<span class="sd"> outputs. The other non-optional parameters are</span>
<span class="sd"> self-explanatory. If ``test_data`` is provided then the</span>
<span class="sd"> network will be evaluated against the test data after each</span>
<span class="sd"> epoch, and partial progress printed out. This is useful for</span>
<span class="sd"> tracking progress, but slows things down substantially."""</span>
<span class="k">if</span> <span class="n">test_data</span><span class="p">:</span> <span class="n">n_test</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="n">test_data</span><span class="p">)</span>
<span class="n">n</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="n">training_data</span><span class="p">)</span>
<span class="k">for</span> <span class="n">j</span> <span class="ow">in</span> <span class="nb">xrange</span><span class="p">(</span><span class="n">epochs</span><span class="p">):</span>
<span class="n">random</span><span class="o">.</span><span class="n">shuffle</span><span class="p">(</span><span class="n">training_data</span><span class="p">)</span>
<span class="n">mini_batches</span> <span class="o">=</span> <span class="p">[</span>
<span class="n">training_data</span><span class="p">[</span><span class="n">k</span><span class="p">:</span><span class="n">k</span><span class="o">+</span><span class="n">mini_batch_size</span><span class="p">]</span>
<span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="nb">xrange</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">n</span><span class="p">,</span> <span class="n">mini_batch_size</span><span class="p">)]</span>
<span class="k">for</span> <span class="n">mini_batch</span> <span class="ow">in</span> <span class="n">mini_batches</span><span class="p">:</span>
<span class="bp">self</span><span class="o">.</span><span class="n">update_mini_batch</span><span class="p">(</span><span class="n">mini_batch</span><span class="p">,</span> <span class="n">eta</span><span class="p">)</span>
<span class="k">if</span> <span class="n">test_data</span><span class="p">:</span>
<span class="k">print</span> <span class="s2">"Epoch {0}: {1} / {2}"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span>
<span class="n">j</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">evaluate</span><span class="p">(</span><span class="n">test_data</span><span class="p">),</span> <span class="n">n_test</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">print</span> <span class="s2">"Epoch {0} complete"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">j</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">update_mini_batch</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">mini_batch</span><span class="p">,</span> <span class="n">eta</span><span class="p">):</span>
<span class="sd">"""Update the network's weights and biases by applying</span>
<span class="sd"> gradient descent using backpropagation to a single mini batch.</span>
<span class="sd"> The ``mini_batch`` is a list of tuples ``(x, y)``, and ``eta``</span>
<span class="sd"> is the learning rate."""</span>
<span class="n">nabla_b</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">b</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="k">for</span> <span class="n">b</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">]</span>
<span class="n">nabla_w</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">w</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="k">for</span> <span class="n">w</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">]</span>
<span class="k">for</span> <span class="n">x</span><span class="p">,</span> <span class="n">y</span> <span class="ow">in</span> <span class="n">mini_batch</span><span class="p">:</span>
<span class="n">delta_nabla_b</span><span class="p">,</span> <span class="n">delta_nabla_w</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">backprop</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span>
<span class="n">nabla_b</span> <span class="o">=</span> <span class="p">[</span><span class="n">nb</span><span class="o">+</span><span class="n">dnb</span> <span class="k">for</span> <span class="n">nb</span><span class="p">,</span> <span class="n">dnb</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">nabla_b</span><span class="p">,</span> <span class="n">delta_nabla_b</span><span class="p">)]</span>
<span class="n">nabla_w</span> <span class="o">=</span> <span class="p">[</span><span class="n">nw</span><span class="o">+</span><span class="n">dnw</span> <span class="k">for</span> <span class="n">nw</span><span class="p">,</span> <span class="n">dnw</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">nabla_w</span><span class="p">,</span> <span class="n">delta_nabla_w</span><span class="p">)]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">weights</span> <span class="o">=</span> <span class="p">[</span><span class="n">w</span><span class="o">-</span><span class="p">(</span><span class="n">eta</span><span class="o">/</span><span class="nb">len</span><span class="p">(</span><span class="n">mini_batch</span><span class="p">))</span><span class="o">*</span><span class="n">nw</span>
<span class="k">for</span> <span class="n">w</span><span class="p">,</span> <span class="n">nw</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">,</span> <span class="n">nabla_w</span><span class="p">)]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">biases</span> <span class="o">=</span> <span class="p">[</span><span class="n">b</span><span class="o">-</span><span class="p">(</span><span class="n">eta</span><span class="o">/</span><span class="nb">len</span><span class="p">(</span><span class="n">mini_batch</span><span class="p">))</span><span class="o">*</span><span class="n">nb</span>
<span class="k">for</span> <span class="n">b</span><span class="p">,</span> <span class="n">nb</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">,</span> <span class="n">nabla_b</span><span class="p">)]</span>
<span class="k">def</span> <span class="nf">backprop</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="sd">"""Return a tuple ``(nabla_b, nabla_w)`` representing the</span>
<span class="sd"> gradient for the cost function C_x. ``nabla_b`` and</span>
<span class="sd"> ``nabla_w`` are layer-by-layer lists of numpy arrays, similar</span>
<span class="sd"> to ``self.biases`` and ``self.weights``."""</span>
<span class="n">nabla_b</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">b</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="k">for</span> <span class="n">b</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">]</span>
<span class="n">nabla_w</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">w</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="k">for</span> <span class="n">w</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">]</span>
<span class="c1"># feedforward</span>
<span class="n">activation</span> <span class="o">=</span> <span class="n">x</span>
<span class="n">activations</span> <span class="o">=</span> <span class="p">[</span><span class="n">x</span><span class="p">]</span> <span class="c1"># list to store all the activations, layer by layer</span>
<span class="n">zs</span> <span class="o">=</span> <span class="p">[]</span> <span class="c1"># list to store all the z vectors, layer by layer</span>
<span class="k">for</span> <span class="n">b</span><span class="p">,</span> <span class="n">w</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">biases</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">):</span>
<span class="n">z</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">activation</span><span class="p">)</span><span class="o">+</span><span class="n">b</span>
<span class="n">zs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">z</span><span class="p">)</span>
<span class="n">activation</span> <span class="o">=</span> <span class="n">sigmoid</span><span class="p">(</span><span class="n">z</span><span class="p">)</span>
<span class="n">activations</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">activation</span><span class="p">)</span>
<span class="c1"># backward pass</span>
<span class="n">delta</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">cost_derivative</span><span class="p">(</span><span class="n">activations</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">],</span> <span class="n">y</span><span class="p">)</span> <span class="o">*</span> \
<span class="n">sigmoid_prime</span><span class="p">(</span><span class="n">zs</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">])</span>
<span class="n">nabla_b</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">=</span> <span class="n">delta</span>
<span class="n">nabla_w</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">delta</span><span class="p">,</span> <span class="n">activations</span><span class="p">[</span><span class="o">-</span><span class="mi">2</span><span class="p">]</span><span class="o">.</span><span class="n">transpose</span><span class="p">())</span>
<span class="c1"># Note that the variable l in the loop below is used a little</span>
<span class="c1"># differently to the notation in Chapter 2 of the book. Here,</span>
<span class="c1"># l = 1 means the last layer of neurons, l = 2 is the</span>
<span class="c1"># second-last layer, and so on. It's a renumbering of the</span>
<span class="c1"># scheme in the book, used here to take advantage of the fact</span>
<span class="c1"># that Python can use negative indices in lists.</span>
<span class="k">for</span> <span class="n">l</span> <span class="ow">in</span> <span class="nb">xrange</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span><span class="p">):</span>
<span class="n">z</span> <span class="o">=</span> <span class="n">zs</span><span class="p">[</span><span class="o">-</span><span class="n">l</span><span class="p">]</span>
<span class="n">sp</span> <span class="o">=</span> <span class="n">sigmoid_prime</span><span class="p">(</span><span class="n">z</span><span class="p">)</span>
<span class="n">delta</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">weights</span><span class="p">[</span><span class="o">-</span><span class="n">l</span><span class="o">+</span><span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">transpose</span><span class="p">(),</span> <span class="n">delta</span><span class="p">)</span> <span class="o">*</span> <span class="n">sp</span>
<span class="n">nabla_b</span><span class="p">[</span><span class="o">-</span><span class="n">l</span><span class="p">]</span> <span class="o">=</span> <span class="n">delta</span>
<span class="n">nabla_w</span><span class="p">[</span><span class="o">-</span><span class="n">l</span><span class="p">]</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">dot</span><span class="p">(</span><span class="n">delta</span><span class="p">,</span> <span class="n">activations</span><span class="p">[</span><span class="o">-</span><span class="n">l</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">transpose</span><span class="p">())</span>
<span class="k">return</span> <span class="p">(</span><span class="n">nabla_b</span><span class="p">,</span> <span class="n">nabla_w</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">evaluate</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">test_data</span><span class="p">):</span>
<span class="sd">"""Return the number of test inputs for which the neural</span>
<span class="sd"> network outputs the correct result. Note that the neural</span>
<span class="sd"> network's output is assumed to be the index of whichever</span>
<span class="sd"> neuron in the final layer has the highest activation."""</span>
<span class="n">test_results</span> <span class="o">=</span> <span class="p">[(</span><span class="n">np</span><span class="o">.</span><span class="n">argmax</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">feedforward</span><span class="p">(</span><span class="n">x</span><span class="p">)),</span> <span class="n">y</span><span class="p">)</span>
<span class="k">for</span> <span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span> <span class="ow">in</span> <span class="n">test_data</span><span class="p">]</span>
<span class="k">return</span> <span class="nb">sum</span><span class="p">(</span><span class="nb">int</span><span class="p">(</span><span class="n">x</span> <span class="o">==</span> <span class="n">y</span><span class="p">)</span> <span class="k">for</span> <span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span> <span class="ow">in</span> <span class="n">test_results</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">cost_derivative</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">output_activations</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="sd">"""Return the vector of partial derivatives \partial C_x /</span>
<span class="sd"> \partial a for the output activations."""</span>
<span class="k">return</span> <span class="p">(</span><span class="n">output_activations</span><span class="o">-</span><span class="n">y</span><span class="p">)</span>
<span class="c1">#### Miscellaneous functions</span>
<span class="k">def</span> <span class="nf">sigmoid</span><span class="p">(</span><span class="n">z</span><span class="p">):</span>
<span class="sd">"""The sigmoid function."""</span>
<span class="k">return</span> <span class="mf">1.0</span><span class="o">/</span><span class="p">(</span><span class="mf">1.0</span><span class="o">+</span><span class="n">np</span><span class="o">.</span><span class="n">exp</span><span class="p">(</span><span class="o">-</span><span class="n">z</span><span class="p">))</span>
<span class="k">def</span> <span class="nf">sigmoid_prime</span><span class="p">(</span><span class="n">z</span><span class="p">):</span>
<span class="sd">"""Derivative of the sigmoid function."""</span>
<span class="k">return</span> <span class="n">sigmoid</span><span class="p">(</span><span class="n">z</span><span class="p">)</span><span class="o">*</span><span class="p">(</span><span class="mi">1</span><span class="o">-</span><span class="n">sigmoid</span><span class="p">(</span><span class="n">z</span><span class="p">))</span>
</pre></div>
</p><p>How well does the program recognize handwritten digits? Well, let'sstart by loading in the MNIST data. I'll do this using a littlehelper program, <tt>mnist_loader.py</tt>, to be described below. Weexecute the following commands in a Python shell,</p><p><div class="highlight"><pre><span></span><span class="o">>>></span> <span class="kn">import</span> <span class="nn">mnist_loader</span>
<span class="o">>>></span> <span class="n">training_data</span><span class="p">,</span> <span class="n">validation_data</span><span class="p">,</span> <span class="n">test_data</span> <span class="o">=</span> \
<span class="o">...</span> <span class="n">mnist_loader</span><span class="o">.</span><span class="n">load_data_wrapper</span><span class="p">()</span>
</pre></div>
</p><p>Of course, this could also be done in a separate Python program, butif you're following along it's probably easiest to do in a Pythonshell. </p><p>After loading the MNIST data, we'll set up a <tt>Network</tt> with $30$hidden neurons. We do this after importing the Python program listedabove, which is named <tt>network</tt>,</p><p><div class="highlight"><pre><span></span><span class="o">>>></span> <span class="kn">import</span> <span class="nn">network</span>
<span class="o">>>></span> <span class="n">net</span> <span class="o">=</span> <span class="n">network</span><span class="o">.</span><span class="n">Network</span><span class="p">([</span><span class="mi">784</span><span class="p">,</span> <span class="mi">30</span><span class="p">,</span> <span class="mi">10</span><span class="p">])</span>
</pre></div>
</p><p>Finally, we'll use stochastic gradient descent to learn from the MNIST<tt>training_data</tt> over 30 epochs, with a mini-batch size of 10, and alearning rate of $\eta = 3.0$, </p><p><div class="highlight"><pre><span></span><span class="o">>>></span> <span class="n">net</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">training_data</span><span class="p">,</span> <span class="mi">30</span><span class="p">,</span> <span class="mi">10</span><span class="p">,</span> <span class="mf">3.0</span><span class="p">,</span> <span class="n">test_data</span><span class="o">=</span><span class="n">test_data</span><span class="p">)</span>
</pre></div>
</p><p>Note that if you're running the code as you read along, it will takesome time to execute - for a typical machine (as of 2015) it willlikely take a few minutes to run. I suggest you set things running,continue to read, and periodically check the output from the code. Ifyou're in a rush you can speed things up by decreasing the number ofepochs, by decreasing the number of hidden neurons, or by using onlypart of the training data. Note that production code would be much,much faster: these Python scripts are intended to help you understandhow neural nets work, not to be high-performance code! And, ofcourse, once we've trained a network it can be run very quicklyindeed, on almost any computing platform. For example, once we'velearned a good set of weights and biases for a network, it can easilybe ported to run in Javascript in a web browser, or as a native app ona mobile device. In any case, here is a partial transcript of theoutput of one training run of the neural network. The transcriptshows the number of test images correctly recognized by the neuralnetwork after each epoch of training. As you can see, after just asingle epoch this has reached 9,129 out of 10,000, and the numbercontinues to grow,</p><p><div class="highlight"><pre><span></span>Epoch 0: 9129 / 10000
Epoch 1: 9295 / 10000
Epoch 2: 9348 / 10000
...
Epoch 27: 9528 / 10000
Epoch 28: 9542 / 10000
Epoch 29: 9534 / 10000
</pre></div>
</p><p>That is, the trained network gives us a classification rate of about$95$ percent - $95.42$ percent at its peak ("Epoch 28")! That'squite encouraging as a first attempt. I should warn you, however,that if you run the code then your results are not necessarily goingto be quite the same as mine, since we'll be initializing our networkusing (different) random weights and biases. To generate results inthis chapter I've taken best-of-three runs.</p><p>Let's rerun the above experiment, changing the number of hiddenneurons to $100$. As was the case earlier, if you're running the codeas you read along, you should be warned that it takes quite a while toexecute (on my machine this experiment takes tens of seconds for eachtraining epoch), so it's wise to continue reading in parallel whilethe code executes.</p><p><div class="highlight"><pre><span></span><span class="o">>>></span> <span class="n">net</span> <span class="o">=</span> <span class="n">network</span><span class="o">.</span><span class="n">Network</span><span class="p">([</span><span class="mi">784</span><span class="p">,</span> <span class="mi">100</span><span class="p">,</span> <span class="mi">10</span><span class="p">])</span>
<span class="o">>>></span> <span class="n">net</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">training_data</span><span class="p">,</span> <span class="mi">30</span><span class="p">,</span> <span class="mi">10</span><span class="p">,</span> <span class="mf">3.0</span><span class="p">,</span> <span class="n">test_data</span><span class="o">=</span><span class="n">test_data</span><span class="p">)</span>
</pre></div>
</p><p>Sure enough, this improves the results to $96.59$ percent. At leastin this case, using more hidden neurons helps us get betterresults*<span class="marginnote">*Reader feedback indicates quite some variation in results for this experiment, and some training runs give results quite a bit worse. Using the techniques introduced in chapter 3 will greatly reduce the variation in performance across different training runs for our networks.</span>.</p><p>Of course, to obtain these accuracies I had to make specific choicesfor the number of epochs of training, the mini-batch size, and thelearning rate, $\eta$. As I mentioned above, these are known ashyper-parameters for our neural network, in order to distinguish themfrom the parameters (weights and biases) learnt by our learningalgorithm. If we choose our hyper-parameters poorly, we can get badresults. Suppose, for example, that we'd chosen the learning rate tobe $\eta = 0.001$,</p><p><div class="highlight"><pre><span></span><span class="o">>>></span> <span class="n">net</span> <span class="o">=</span> <span class="n">network</span><span class="o">.</span><span class="n">Network</span><span class="p">([</span><span class="mi">784</span><span class="p">,</span> <span class="mi">100</span><span class="p">,</span> <span class="mi">10</span><span class="p">])</span>
<span class="o">>>></span> <span class="n">net</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">training_data</span><span class="p">,</span> <span class="mi">30</span><span class="p">,</span> <span class="mi">10</span><span class="p">,</span> <span class="mf">0.001</span><span class="p">,</span> <span class="n">test_data</span><span class="o">=</span><span class="n">test_data</span><span class="p">)</span>
</pre></div>
</p><p>The results are much less encouraging,<div class="highlight"><pre><span></span>Epoch 0: 1139 / 10000
Epoch 1: 1136 / 10000
Epoch 2: 1135 / 10000
...
Epoch 27: 2101 / 10000
Epoch 28: 2123 / 10000
Epoch 29: 2142 / 10000
</pre></div>
However, you can see that the performance of the network is gettingslowly better over time. That suggests increasing the learning rate,say to $\eta = 0.01$. If we do that, we get better results, whichsuggests increasing the learning rate again. (If making a changeimproves things, try doing more!) If we do that several times over,we'll end up with a learning rate of something like $\eta = 1.0$ (andperhaps fine tune to $3.0$), which is close to our earlierexperiments. So even though we initially made a poor choice ofhyper-parameters, we at least got enough information to help usimprove our choice of hyper-parameters.</p><p>In general, debugging a neural network can be challenging. This isespecially true when the initial choice of hyper-parameters producesresults no better than random noise. Suppose we try the successful 30hidden neuron network architecture from earlier, but with the learningrate changed to $\eta = 100.0$:<div class="highlight"><pre><span></span><span class="o">>>></span> <span class="n">net</span> <span class="o">=</span> <span class="n">network</span><span class="o">.</span><span class="n">Network</span><span class="p">([</span><span class="mi">784</span><span class="p">,</span> <span class="mi">30</span><span class="p">,</span> <span class="mi">10</span><span class="p">])</span>
<span class="o">>>></span> <span class="n">net</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">training_data</span><span class="p">,</span> <span class="mi">30</span><span class="p">,</span> <span class="mi">10</span><span class="p">,</span> <span class="mf">100.0</span><span class="p">,</span> <span class="n">test_data</span><span class="o">=</span><span class="n">test_data</span><span class="p">)</span>
</pre></div>
At this point we've actually gone too far, and the learning rate istoo high:<div class="highlight"><pre><span></span><span class="n">Epoch</span> <span class="mi">0</span><span class="p">:</span> <span class="mi">1009</span> <span class="o">/</span> <span class="mi">10000</span>
<span class="n">Epoch</span> <span class="mi">1</span><span class="p">:</span> <span class="mi">1009</span> <span class="o">/</span> <span class="mi">10000</span>
<span class="n">Epoch</span> <span class="mi">2</span><span class="p">:</span> <span class="mi">1009</span> <span class="o">/</span> <span class="mi">10000</span>
<span class="n">Epoch</span> <span class="mi">3</span><span class="p">:</span> <span class="mi">1009</span> <span class="o">/</span> <span class="mi">10000</span>
<span class="o">...</span>
<span class="n">Epoch</span> <span class="mi">27</span><span class="p">:</span> <span class="mi">982</span> <span class="o">/</span> <span class="mi">10000</span>
<span class="n">Epoch</span> <span class="mi">28</span><span class="p">:</span> <span class="mi">982</span> <span class="o">/</span> <span class="mi">10000</span>
<span class="n">Epoch</span> <span class="mi">29</span><span class="p">:</span> <span class="mi">982</span> <span class="o">/</span> <span class="mi">10000</span>
</pre></div>
Now imagine that we were coming to this problem for the first time.Of course, we <em>know</em> from our earlier experiments that the rightthing to do is to decrease the learning rate. But if we were comingto this problem for the first time then there wouldn't be much in theoutput to guide us on what to do. We might worry not only about thelearning rate, but about every other aspect of our neural network. Wemight wonder if we've initialized the weights and biases in a way thatmakes it hard for the network to learn? Or maybe we don't have enoughtraining data to get meaningful learning? Perhaps we haven't run forenough epochs? Or maybe it's impossible for a neural network withthis architecture to learn to recognize handwritten digits? Maybe thelearning rate is too <em>low</em>? Or, maybe, the learning rate is toohigh? When you're coming to a problem for the first time, you're notalways sure.</p><p>The lesson to take away from this is that debugging a neural networkis not trivial, and, just as for ordinary programming, there is an artto it. You need to learn that art of debugging in order to get goodresults from neural networks. More generally, we need to developheuristics for choosing good hyper-parameters and a good architecture.We'll discuss all these at length through the book, including how Ichose the hyper-parameters above.</p><p><h4><a name="exercise_420023"></a><a href="#exercise_420023">Exercise</a></h4><ul></p><p><li> Try creating a network with just two layers - an input and an output layer, no hidden layer - with 784 and 10 neurons, respectively. Train the network using stochastic gradient descent. What classification accuracy can you achieve?</ul></p><p></p><p>Earlier, I skipped over the details of how the MNIST data is loaded.It's pretty straightforward. For completeness, here's the code. Thedata structures used to store the MNIST data are described in thedocumentation strings - it's straightforward stuff, tuples and listsof Numpy <tt>ndarray</tt> objects (think of them as vectors if you'renot familiar with <tt>ndarray</tt>s):</p><p><div class="highlight"><pre><span></span><span class="sd">"""</span>
<span class="sd">mnist_loader</span>
<span class="sd">~~~~~~~~~~~~</span>
<span class="sd">A library to load the MNIST image data. For details of the data</span>
<span class="sd">structures that are returned, see the doc strings for ``load_data``</span>
<span class="sd">and ``load_data_wrapper``. In practice, ``load_data_wrapper`` is the</span>
<span class="sd">function usually called by our neural network code.</span>
<span class="sd">"""</span>
<span class="c1">#### Libraries</span>
<span class="c1"># Standard library</span>
<span class="kn">import</span> <span class="nn">cPickle</span>
<span class="kn">import</span> <span class="nn">gzip</span>
<span class="c1"># Third-party libraries</span>
<span class="kn">import</span> <span class="nn">numpy</span> <span class="kn">as</span> <span class="nn">np</span>
<span class="k">def</span> <span class="nf">load_data</span><span class="p">():</span>
<span class="sd">"""Return the MNIST data as a tuple containing the training data,</span>
<span class="sd"> the validation data, and the test data.</span>
<span class="sd"> The ``training_data`` is returned as a tuple with two entries.</span>
<span class="sd"> The first entry contains the actual training images. This is a</span>
<span class="sd"> numpy ndarray with 50,000 entries. Each entry is, in turn, a</span>
<span class="sd"> numpy ndarray with 784 values, representing the 28 * 28 = 784</span>
<span class="sd"> pixels in a single MNIST image.</span>
<span class="sd"> The second entry in the ``training_data`` tuple is a numpy ndarray</span>
<span class="sd"> containing 50,000 entries. Those entries are just the digit</span>
<span class="sd"> values (0...9) for the corresponding images contained in the first</span>
<span class="sd"> entry of the tuple.</span>
<span class="sd"> The ``validation_data`` and ``test_data`` are similar, except</span>
<span class="sd"> each contains only 10,000 images.</span>
<span class="sd"> This is a nice data format, but for use in neural networks it's</span>
<span class="sd"> helpful to modify the format of the ``training_data`` a little.</span>
<span class="sd"> That's done in the wrapper function ``load_data_wrapper()``, see</span>
<span class="sd"> below.</span>
<span class="sd"> """</span>
<span class="n">f</span> <span class="o">=</span> <span class="n">gzip</span><span class="o">.</span><span class="n">open</span><span class="p">(</span><span class="s1">'../data/mnist.pkl.gz'</span><span class="p">,</span> <span class="s1">'rb'</span><span class="p">)</span>
<span class="n">training_data</span><span class="p">,</span> <span class="n">validation_data</span><span class="p">,</span> <span class="n">test_data</span> <span class="o">=</span> <span class="n">cPickle</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">f</span><span class="p">)</span>
<span class="n">f</span><span class="o">.</span><span class="n">close</span><span class="p">()</span>
<span class="k">return</span> <span class="p">(</span><span class="n">training_data</span><span class="p">,</span> <span class="n">validation_data</span><span class="p">,</span> <span class="n">test_data</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">load_data_wrapper</span><span class="p">():</span>
<span class="sd">"""Return a tuple containing ``(training_data, validation_data,</span>
<span class="sd"> test_data)``. Based on ``load_data``, but the format is more</span>
<span class="sd"> convenient for use in our implementation of neural networks.</span>
<span class="sd"> In particular, ``training_data`` is a list containing 50,000</span>
<span class="sd"> 2-tuples ``(x, y)``. ``x`` is a 784-dimensional numpy.ndarray</span>
<span class="sd"> containing the input image. ``y`` is a 10-dimensional</span>
<span class="sd"> numpy.ndarray representing the unit vector corresponding to the</span>
<span class="sd"> correct digit for ``x``.</span>
<span class="sd"> ``validation_data`` and ``test_data`` are lists containing 10,000</span>
<span class="sd"> 2-tuples ``(x, y)``. In each case, ``x`` is a 784-dimensional</span>
<span class="sd"> numpy.ndarry containing the input image, and ``y`` is the</span>
<span class="sd"> corresponding classification, i.e., the digit values (integers)</span>
<span class="sd"> corresponding to ``x``.</span>
<span class="sd"> Obviously, this means we're using slightly different formats for</span>
<span class="sd"> the training data and the validation / test data. These formats</span>
<span class="sd"> turn out to be the most convenient for use in our neural network</span>
<span class="sd"> code."""</span>
<span class="n">tr_d</span><span class="p">,</span> <span class="n">va_d</span><span class="p">,</span> <span class="n">te_d</span> <span class="o">=</span> <span class="n">load_data</span><span class="p">()</span>
<span class="n">training_inputs</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="p">(</span><span class="mi">784</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span> <span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="n">tr_d</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="n">training_results</span> <span class="o">=</span> <span class="p">[</span><span class="n">vectorized_result</span><span class="p">(</span><span class="n">y</span><span class="p">)</span> <span class="k">for</span> <span class="n">y</span> <span class="ow">in</span> <span class="n">tr_d</span><span class="p">[</span><span class="mi">1</span><span class="p">]]</span>
<span class="n">training_data</span> <span class="o">=</span> <span class="nb">zip</span><span class="p">(</span><span class="n">training_inputs</span><span class="p">,</span> <span class="n">training_results</span><span class="p">)</span>
<span class="n">validation_inputs</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="p">(</span><span class="mi">784</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span> <span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="n">va_d</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="n">validation_data</span> <span class="o">=</span> <span class="nb">zip</span><span class="p">(</span><span class="n">validation_inputs</span><span class="p">,</span> <span class="n">va_d</span><span class="p">[</span><span class="mi">1</span><span class="p">])</span>
<span class="n">test_inputs</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="p">(</span><span class="mi">784</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span> <span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="n">te_d</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="n">test_data</span> <span class="o">=</span> <span class="nb">zip</span><span class="p">(</span><span class="n">test_inputs</span><span class="p">,</span> <span class="n">te_d</span><span class="p">[</span><span class="mi">1</span><span class="p">])</span>
<span class="k">return</span> <span class="p">(</span><span class="n">training_data</span><span class="p">,</span> <span class="n">validation_data</span><span class="p">,</span> <span class="n">test_data</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">vectorized_result</span><span class="p">(</span><span class="n">j</span><span class="p">):</span>
<span class="sd">"""Return a 10-dimensional unit vector with a 1.0 in the jth</span>
<span class="sd"> position and zeroes elsewhere. This is used to convert a digit</span>
<span class="sd"> (0...9) into a corresponding desired output from the neural</span>
<span class="sd"> network."""</span>
<span class="n">e</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">((</span><span class="mi">10</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span>
<span class="n">e</span><span class="p">[</span><span class="n">j</span><span class="p">]</span> <span class="o">=</span> <span class="mf">1.0</span>
<span class="k">return</span> <span class="n">e</span>
</pre></div>
</p><p>I said above that our program gets pretty good results. What doesthat mean? Good compared to what? It's informative to have somesimple (non-neural-network) baseline tests to compare against, tounderstand what it means to perform well. The simplest baseline ofall, of course, is to randomly guess the digit. That'll be rightabout ten percent of the time. We're doing much better than that!</p><p>What about a less trivial baseline? Let's try an extremely simpleidea: we'll look at how <em>dark</em> an image is. For instance, animage of a $2$ will typically be quite a bit darker than an image of a$1$, just because more pixels are blackened out, as the followingexamples illustrate:</p><p><center><img src="images/mnist_2_and_1.png" width="256px"></center></p><p>This suggests using the training data to compute average darknessesfor each digit, $0, 1, 2,\ldots, 9$. When presented with a new image,we compute how dark the image is, and then guess that it's whicheverdigit has the closest average darkness. This is a simple procedure,and is easy to code up, so I won't explicitly write out the code -if you're interested it's in the<a href="https://github.com/mnielsen/neural-networks-and-deep-learning/blob/master/src/mnist_average_darkness.py">GitHub repository</a>. But it's a big improvement over random guessing,getting $2,225$ of the $10,000$ test images correct, i.e., $22.25$percent accuracy.</p><p><a name="SVM"></a></p><p>It's not difficult to find other ideas which achieve accuracies in the$20$ to $50$ percent range. If you work a bit harder you can get upover $50$ percent. But to get much higher accuracies it helps to useestablished machine learning algorithms. Let's try using one of thebest known algorithms, the <em>support vector machine</em>or <em>SVM</em>. If you're notfamiliar with SVMs, not to worry, we're not going to need tounderstand the details of how SVMs work. Instead, we'll use a Pythonlibrary called<a href="http://scikit-learn.org/stable/">scikit-learn</a>,which provides a simple Python interface to a fast C-based library forSVMs known as<a href="http://www.csie.ntu.edu.tw/~cjlin/libsvm/">LIBSVM</a>.</p><p>If we run scikit-learn's SVM classifier using the default settings,then it gets 9,435 of 10,000 test images correct. (The code isavailable<a href="https://github.com/mnielsen/neural-networks-and-deep-learning/blob/master/src/mnist_svm.py">here</a>.)That's a big improvement over our naive approach of classifying animage based on how dark it is. Indeed, it means that the SVM isperforming roughly as well as our neural networks, just a littleworse. In later chapters we'll introduce new techniques that enableus to improve our neural networks so that they perform much betterthan the SVM.</p><p>That's not the end of the story, however. The 9,435 of 10,000 resultis for scikit-learn's default settings for SVMs. SVMs have a numberof tunable parameters, and it's possible to search for parameterswhich improve this out-of-the-box performance. I won't explicitly dothis search, but instead refer you to<a href="http://peekaboo-vision.blogspot.de/2010/09/mnist-for-ever.html">this blog post</a> by <a href="http://peekaboo-vision.blogspot.ca/">Andreas Mueller</a> if you'd like to know more. Mueller shows that with somework optimizing the SVM's parameters it's possible to get theperformance up above 98.5 percent accuracy. In other words, awell-tuned SVM only makes an error on about one digit in 70. That'spretty good! Can neural networks do better?</p><p>In fact, they can. At present, well-designed neural networksoutperform every other technique for solving MNIST, including SVMs.The current (2013) record is classifying 9,979 of 10,000 imagescorrectly. This was done by <a href="http://www.cs.nyu.edu/~wanli/">Li Wan</a>, <a href="http://www.matthewzeiler.com/">Matthew Zeiler</a>, SixinZhang, <a href="http://yann.lecun.com/">Yann LeCun</a>, and<a href="http://cs.nyu.edu/~fergus/pmwiki/pmwiki.php">Rob Fergus</a>.We'll see most of the techniques they used later in the book. At thatlevel the performance is close to human-equivalent, and is arguablybetter, since quite a few of the MNIST images are difficult even forhumans to recognize with confidence, for example:</p><p><center><img src="images/mnist_really_bad_images.png" width="560px"></center></p><p>I trust you'll agree that those are tough to classify! With imageslike these in the MNIST data set it's remarkable that neural networkscan accurately classify all but 21 of the 10,000 test images.Usually, when programming we believe that solving a complicatedproblem like recognizing the MNIST digits requires a sophisticatedalgorithm. But even the neural networks in the Wan <em>et al</em> paperjust mentioned involve quite simple algorithms, variations on thealgorithm we've seen in this chapter. All the complexity is learned,automatically, from the training data. In some sense, the moral ofboth our results and those in more sophisticated papers, is that forsome problems:<center> sophisticated algorithm $\leq$ simple learning algorithm + good training data.</center></p><p><h3><a name="toward_deep_learning"></a><a href="#toward_deep_learning">Toward deep learning</a></h3></p><p>While our neural network gives impressive performance, thatperformance is somewhat mysterious. The weights and biases in thenetwork were discovered automatically. And that means we don'timmediately have an explanation of how the network does what it does.Can we find some way to understand the principles by which our networkis classifying handwritten digits? And, given such principles, can wedo better?</p><p>To put these questions more starkly, suppose that a few decades henceneural networks lead to artificial intelligence (AI). Will weunderstand how such intelligent networks work? Perhaps the networkswill be opaque to us, with weights and biases we don't understand,because they've been learned automatically. In the early days of AIresearch people hoped that the effort to build an AI would also helpus understand the principles behind intelligence and, maybe, thefunctioning of the human brain. But perhaps the outcome will be thatwe end up understanding neither the brain nor how artificialintelligence works!</p><p>To address these questions, let's think back to the interpretation ofartificial neurons that I gave at the start of the chapter, as a meansof weighing evidence. Suppose we want to determine whether an imageshows a human face or not:</p><p> </p><p> <span class="marginnote">Credits: 1. <a href="http://commons.wikimedia.org/wiki/User:ST">Ester Inbar</a>. 2. Unknown. 3. NASA, ESA, G. Illingworth, D. Magee, and P. Oesch (University of California, Santa Cruz), R. Bouwens (Leiden University), and the HUDF09 Team. Click on the images for more details.</span></p><p> <a href="http://commons.wikimedia.org/wiki/File:Kangaroo_ST_03.JPG"><img src="images/Kangaroo.JPG" height="190px"/></a> <a href="http://commons.wikimedia.org/wiki/File:Albert_Einstein_at_the_age_of_three_(1882).jpg"><img src="images/Einstein_crop.jpg" height="190px"/></a> <a href="http://commons.wikimedia.org/wiki/File:The_Hubble_eXtreme_Deep_Field.jpg"><img src="images/hubble.jpg" height="190px"/></a> </p><p>We could attack this problem the same way we attacked handwritingrecognition - by using the pixels in the image as input to a neuralnetwork, with the output from the network a single neuron indicatingeither "Yes, it's a face" or "No, it's not a face".</p><p>Let's suppose we do this, but that we're not using a learningalgorithm. Instead, we're going to try to design a network by hand,choosing appropriate weights and biases. How might we go about it?Forgetting neural networks entirely for the moment, a heuristic wecould use is to decompose the problem into sub-problems: does theimage have an eye in the top left? Does it have an eye in the topright? Does it have a nose in the middle? Does it have a mouth inthe bottom middle? Is there hair on top? And so on.</p><p>If the answers to several of these questions are "yes", or even just"probably yes", then we'd conclude that the image is likely to be aface. Conversely, if the answers to most of the questions are "no",then the image probably isn't a face.</p><p>Of course, this is just a rough heuristic, and it suffers from manydeficiencies. Maybe the person is bald, so they have no hair. Maybewe can only see part of the face, or the face is at an angle, so someof the facial features are obscured. Still, the heuristic suggeststhat if we can solve the sub-problems using neural networks, thenperhaps we can build a neural network for face-detection, by combiningthe networks for the sub-problems. Here's a possible architecture,with rectangles denoting the sub-networks. Note that this isn'tintended as a realistic approach to solving the face-detectionproblem; rather, it's to help us build intuition about how networksfunction. Here's the architecture:</p><p><center><img src="images/tikz14.png"/></center></p><p>It's also plausible that the sub-networks can be decomposed. Supposewe're considering the question: "Is there an eye in the top left?"This can be decomposed into questions such as: "Is there aneyebrow?"; "Are there eyelashes?"; "Is there an iris?"; and soon. Of course, these questions should really include positionalinformation, as well - "Is the eyebrow in the top left, and abovethe iris?", that kind of thing - but let's keep it simple. Thenetwork to answer the question "Is there an eye in the top left?"can now be decomposed:</p><p><center><img src="images/tikz15.png"/></center></p><p>Those questions too can be broken down, further and further throughmultiple layers. Ultimately, we'll be working with sub-networks thatanswer questions so simple they can easily be answered at the level ofsingle pixels. Those questions might, for example, be about thepresence or absence of very simple shapes at particular points in theimage. Such questions can be answered by single neurons connected tothe raw pixels in the image.</p><p>The end result is a network which breaks down a very complicatedquestion - does this image show a face or not - into very simplequestions answerable at the level of single pixels. It does thisthrough a series of many layers, with early layers answering verysimple and specific questions about the input image, and later layersbuilding up a hierarchy of ever more complex and abstract concepts.Networks with this kind of many-layer structure - two or more hiddenlayers - are called <em>deep neural networks</em>.</p><p></p><p></p><p>Of course, I haven't said how to do this recursive decomposition intosub-networks. It certainly isn't practical to hand-design the weightsand biases in the network. Instead, we'd like to use learningalgorithms so that the network can automatically learn the weights andbiases - and thus, the hierarchy of concepts - from training data.Researchers in the 1980s and 1990s tried using stochastic gradientdescent and backpropagation to train deep networks. Unfortunately,except for a few special architectures, they didn't have much luck.The networks would learn, but very slowly, and in practice often tooslowly to be useful.</p><p>Since 2006, a set of techniques has been developed that enablelearning in deep neural nets. These deep learning techniques arebased on stochastic gradient descent and backpropagation, but alsointroduce new ideas. These techniques have enabled much deeper (andlarger) networks to be trained - people now routinely train networkswith 5 to 10 hidden layers. And, it turns out that these perform farbetter on many problems than shallow neural networks, i.e., networkswith just a single hidden layer. The reason, of course, is theability of deep nets to build up a complex hierarchy of concepts.It's a bit like the way conventional programming languages use modulardesign and ideas about abstraction to enable the creation of complexcomputer programs. Comparing a deep network to a shallow network is abit like comparing a programming language with the ability to makefunction calls to a stripped down language with no ability to makesuch calls. Abstraction takes a different form in neural networksthan it does in conventional programming, but it's just as important.</p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></p><p></div><div class="footer"> <span class="left_footer"> In academic work,
please cite this book as: Michael A. Nielsen, "Neural Networks and
Deep Learning", Determination Press, 2015
<br/>
<br/>
This work is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-nc/3.0/deed.en_GB"
style="color: #eee;">Creative Commons Attribution-NonCommercial 3.0
Unported License</a>. This means you're free to copy, share, and
build on this book, but not to sell it. If you're interested in
commercial use, please <a
href="mailto:mn@michaelnielsen.org">contact me</a>.
</span>
<span class="right_footer">
Last update: Thu Dec 26 15:26:33 2019
<br/>
<br/>
<br/>
<a rel="license" href="http://creativecommons.org/licenses/by-nc/3.0/deed.en_GB"><img alt="Creative Commons Licence" style="border-width:0" src="http://i.creativecommons.org/l/by-nc/3.0/88x31.png" /></a>
</span>
</div>
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','//www.google-analytics.com/analytics.js','ga');
ga('create', 'UA-44208967-1', 'neuralnetworksanddeeplearning.com');
ga('send', 'pageview');
</script>
</body>
</html>