forked from emer/leabra
-
Notifications
You must be signed in to change notification settings - Fork 0
/
layer.go
609 lines (564 loc) · 17.3 KB
/
layer.go
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
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
// Copyright (c) 2019, The Emergent Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package deep
import (
"fmt"
"github.com/chewxy/math32"
"github.com/emer/emergent/emer"
"github.com/emer/leabra/leabra"
"github.com/goki/ki/kit"
)
// deep.Layer is the DeepLeabra layer, based on basic rate-coded leabra.Layer
type Layer struct {
leabra.Layer // access as .Layer
DeepBurst DeepBurstParams `desc:"parameters for computing DeepBurst from act, in Superficial layers (but also needed in Deep layers for deep self connections)"`
DeepCtxt DeepCtxtParams `desc:"parameters for computing DeepCtxt in Deep layers, from BurstCtxt inputs from Super senders"`
DeepTRC DeepTRCParams `desc:"parameters for computing TRC plus-phase (outcome) activations based on TRCBurstGe excitatory input from BurstTRC projections"`
DeepAttn DeepAttnParams `desc:"parameters for computing DeepAttn and DeepLrn attentional modulation signals based on DeepAttn projection inputs integrated into AttnGe excitatory conductances"`
DeepNeurs []Neuron `desc:"slice of extra deep.Neuron state for this layer -- flat list of len = Shape.Len(). You must iterate over index and use pointer to modify values."`
DeepPools []Pool `desc:"extra layer and sub-pool (unit group) statistics used in DeepLeabra -- flat list has at least of 1 for layer, and one for each sub-pool (unit group) if shape supports that (4D). You must iterate over index and use pointer to modify values."`
}
// AsLeabra returns this layer as a leabra.Layer -- all derived layers must redefine
// this to return the base Layer type, so that the LeabraLayer interface does not
// need to include accessors to all the basic stuff
func (ly *Layer) AsLeabra() *leabra.Layer {
return &ly.Layer
}
// AsDeep returns this layer as a deep.Layer -- all derived layers must redefine
// this to return the deep Layer type, so that the DeepLayer interface does not
// need to include accessors to all the fields.
func (ly *Layer) AsDeep() *Layer {
return ly
}
func (ly *Layer) Defaults() {
ly.Layer.Defaults()
ly.DeepBurst.Defaults()
ly.DeepCtxt.Defaults()
ly.DeepTRC.Defaults()
ly.DeepAttn.Defaults()
}
// UpdateParams updates all params given any changes that might have been made to individual values
// including those in the receiving projections of this layer
func (ly *Layer) UpdateParams() {
ly.Layer.UpdateParams()
ly.DeepBurst.Update()
ly.DeepCtxt.Update()
ly.DeepTRC.Update()
ly.DeepAttn.Update()
}
// UnitVarNames returns a list of variable names available on the units in this layer
func (ly *Layer) UnitVarNames() []string {
return AllNeuronVars
}
// UnitValsTry is emer.Layer interface method to return values of given variable
func (ly *Layer) UnitValsTry(varNm string) ([]float32, error) {
vidx, err := leabra.NeuronVarByName(varNm)
if err == nil {
return ly.Layer.UnitValsTry(varNm)
}
vidx, err = NeuronVarByName(varNm)
if err != nil {
return nil, err
}
vs := make([]float32, len(ly.DeepNeurs))
for i := range ly.DeepNeurs {
dnr := &ly.DeepNeurs[i]
vs[i] = dnr.VarByIndex(vidx)
}
return vs, nil
}
// UnitValTry returns value of given variable name on given unit,
// using shape-based dimensional index
func (ly *Layer) UnitValTry(varNm string, idx []int) (float32, error) {
_, err := leabra.NeuronVarByName(varNm)
if err == nil {
return ly.Layer.UnitValTry(varNm, idx)
}
fidx := ly.Shp.Offset(idx)
nn := len(ly.DeepNeurs)
if fidx < 0 || fidx >= nn {
return 0, fmt.Errorf("Layer UnitVal index: %v out of range, N = %v", fidx, nn)
}
dnr := &ly.DeepNeurs[fidx]
return dnr.VarByName(varNm)
}
// UnitVal1DTry returns value of given variable name on given unit,
// using 1-dimensional index.
func (ly *Layer) UnitVal1DTry(varNm string, idx int) (float32, error) {
_, err := leabra.NeuronVarByName(varNm)
if err == nil {
return ly.Layer.UnitVal1DTry(varNm, idx)
}
nn := len(ly.DeepNeurs)
if idx < 0 || idx >= nn {
return 0, fmt.Errorf("Layer UnitVal1D index: %v out of range, N = %v", idx, nn)
}
dnr := &ly.DeepNeurs[idx]
return dnr.VarByName(varNm)
}
// Build constructs the layer state, including calling Build on the projections
// you MUST have properly configured the Inhib.Pool.On setting by this point
// to properly allocate Pools for the unit groups if necessary.
func (ly *Layer) Build() error {
err := ly.Layer.Build()
if err != nil {
return err
}
ly.DeepNeurs = make([]Neuron, len(ly.Neurons))
ly.DeepPools = make([]Pool, len(ly.Pools))
return nil
}
//////////////////////////////////////////////////////////////////////////////////////
// Init methods
func (ly *Layer) InitActs() {
ly.Layer.InitActs()
for ni := range ly.DeepNeurs {
dnr := &ly.DeepNeurs[ni]
dnr.ActNoAttn = 0
dnr.DeepBurst = 0
dnr.DeepBurstPrv = 0
dnr.DeepCtxt = 0
dnr.TRCBurstGe = 0
dnr.DeepBurstSent = 0
dnr.AttnGe = 0
dnr.DeepAttn = 0
dnr.DeepLrn = 0
}
}
// GScaleFmAvgAct computes the scaling factor for synaptic input conductances G,
// based on sending layer average activation.
// This attempts to automatically adjust for overall differences in raw activity coming into the units
// to achieve a general target of around .5 to 1 for the integrated G values.
// DeepLeabra version separately normalizes the Deep projection types.
func (ly *Layer) GScaleFmAvgAct() {
totGeRel := float32(0)
totGiRel := float32(0)
totTrcRel := float32(0)
totAttnRel := float32(0)
for _, p := range ly.RcvPrjns {
if p.IsOff() {
continue
}
pj := p.(leabra.LeabraPrjn).AsLeabra()
slay := p.SendLay().(leabra.LeabraLayer).AsLeabra()
slpl := slay.Pools[0]
savg := slpl.ActAvg.ActPAvgEff
snu := len(slay.Neurons)
ncon := pj.RConNAvgMax.Avg
pj.GScale = pj.WtScale.FullScale(savg, float32(snu), ncon)
switch pj.Typ {
case emer.Inhib:
totGiRel += pj.WtScale.Rel
case BurstTRC:
totTrcRel += pj.WtScale.Rel
case DeepAttn:
totAttnRel += pj.WtScale.Rel
default:
// note: BurstCtxt is added in here!
totGeRel += pj.WtScale.Rel
}
}
for _, p := range ly.RcvPrjns {
if p.IsOff() {
continue
}
pj := p.(leabra.LeabraPrjn).AsLeabra()
switch pj.Typ {
case emer.Inhib:
if totGiRel > 0 {
pj.GScale /= totGiRel
}
case BurstTRC:
if totTrcRel > 0 {
pj.GScale /= totTrcRel
}
case DeepAttn:
if totAttnRel > 0 {
pj.GScale /= totAttnRel
}
default:
if totGeRel > 0 {
pj.GScale /= totGeRel
}
}
}
}
func (ly *Layer) DecayState(decay float32) {
ly.Layer.DecayState(decay)
for ni := range ly.DeepNeurs {
dnr := &ly.DeepNeurs[ni]
// if dnr.IsOff() { // not worth checking..
// continue
// }
dnr.ActNoAttn -= decay * (dnr.ActNoAttn - ly.Act.Init.Act)
dnr.DeepBurstSent = 0
}
}
//////////////////////////////////////////////////////////////////////////////////////
// Cycle
// SendGDelta sends change in activation since last sent, if above thresholds.
// Deep version sends either to standard Ge or AttnGe for DeepAttn projections.
func (ly *Layer) SendGDelta(ltime *leabra.Time) {
for ni := range ly.Neurons {
nrn := &ly.Neurons[ni]
if nrn.IsOff() {
continue
}
if nrn.Act > ly.Act.OptThresh.Send {
delta := nrn.Act - nrn.ActSent
if math32.Abs(delta) > ly.Act.OptThresh.Delta {
for _, sp := range ly.SndPrjns {
if sp.IsOff() {
continue
}
pj := sp.(DeepPrjn)
ptyp := pj.Type()
if ptyp == BurstCtxt || ptyp == BurstTRC {
continue
}
if ptyp == DeepAttn {
if ly.DeepAttn.On {
pj.SendAttnGeDelta(ni, delta)
}
} else {
pj.SendGDelta(ni, delta)
}
}
nrn.ActSent = nrn.Act
}
} else if nrn.ActSent > ly.Act.OptThresh.Send {
delta := -nrn.ActSent // un-send the last above-threshold activation to get back to 0
for _, sp := range ly.SndPrjns {
if sp.IsOff() {
continue
}
pj := sp.(DeepPrjn)
ptyp := pj.Type()
if ptyp == BurstCtxt || ptyp == BurstTRC {
continue
}
if ptyp == DeepAttn {
if ly.DeepAttn.On {
pj.SendAttnGeDelta(ni, delta)
}
} else {
pj.SendGDelta(ni, delta)
}
}
nrn.ActSent = 0
}
}
}
// GFmInc integrates new synaptic conductances from increments sent during last SendGDelta.
func (ly *Layer) GFmInc(ltime *leabra.Time) {
if ly.Typ == TRC && ly.DeepBurst.IsBurstQtr(ltime.Quarter) {
// note: TRCBurstGe is sent at *end* of previous cycle, after DeepBurst act is computed
lpl := &ly.DeepPools[0]
if lpl.TRCBurstGe.Max > 0.1 { // have some actual input
for ni := range ly.Neurons {
nrn := &ly.Neurons[ni]
if nrn.IsOff() {
continue
}
dnr := &ly.DeepNeurs[ni]
ly.Act.GRawFmInc(nrn) // key to integrate and reset Inc's
geRaw := ly.DeepTRC.BurstGe(dnr.TRCBurstGe)
ly.Act.Dt.GFmRaw(geRaw, &nrn.Ge) // Ge driven exclusively from Burst
}
return
}
}
ly.Layer.GFmInc(ltime) // regular
if ly.DeepAttn.On {
for _, p := range ly.RcvPrjns {
if p.IsOff() {
continue
}
pj := p.(DeepPrjn)
ptyp := pj.Type()
if ptyp != DeepAttn {
continue
}
pj.RecvAttnGeInc()
}
}
}
// AvgMaxGe computes the average and max Ge stats, used in inhibition
// Deep version also computes AttnGe stats
func (ly *Layer) AvgMaxGe(ltime *leabra.Time) {
ly.Layer.AvgMaxGe(ltime)
ly.LeabraLay.(DeepLayer).AvgMaxAttnGe(ltime)
}
// AvgMaxAttnGe computes the average and max AttnGe stats
func (ly *Layer) AvgMaxAttnGe(ltime *leabra.Time) {
for pi := range ly.DeepPools {
pl := &ly.Pools[pi]
dpl := &ly.DeepPools[pi]
dpl.AttnGe.Init()
for ni := pl.StIdx; ni < pl.EdIdx; ni++ {
dnr := &ly.DeepNeurs[ni]
dpl.AttnGe.UpdateVal(dnr.AttnGe, ni)
}
dpl.AttnGe.CalcAvg()
}
}
// ActFmG computes rate-code activation from Ge, Gi, Gl conductances
// and updates learning running-average activations from that Act
func (ly *Layer) ActFmG(ltime *leabra.Time) {
ly.Layer.ActFmG(ltime)
ly.LeabraLay.(DeepLayer).DeepAttnFmG(ltime)
}
// DeepAttnFmG computes DeepAttn and DeepLrn from AttnGe input,
// and then applies the DeepAttn modulation to the Act activation value.
func (ly *Layer) DeepAttnFmG(ltime *leabra.Time) {
lpl := &ly.DeepPools[0]
attnMax := lpl.AttnGe.Max
for ni := range ly.DeepNeurs {
nrn := &ly.Neurons[ni]
if nrn.IsOff() {
continue
}
dnr := &ly.DeepNeurs[ni]
switch {
case !ly.DeepAttn.On:
dnr.DeepAttn = 1
dnr.DeepLrn = 1
case ly.Typ == Deep:
dnr.DeepAttn = nrn.Act // record Deep activation = DeepAttn signal coming from deep layers
dnr.DeepLrn = 1
case ly.Typ == TRC:
dnr.DeepAttn = 1
dnr.DeepLrn = 1
default:
if attnMax < ly.DeepAttn.Thr {
dnr.DeepAttn = 1
dnr.DeepLrn = 1
} else {
dnr.DeepLrn = dnr.AttnGe / attnMax
dnr.DeepAttn = ly.DeepAttn.DeepAttnFmG(dnr.DeepLrn)
}
}
dnr.ActNoAttn = nrn.Act
nrn.Act *= dnr.DeepAttn
}
}
// AvgMaxAct computes the average and max Act stats, used in inhibition
// Deep version also computes AvgMaxActNoAttn
func (ly *Layer) AvgMaxAct(ltime *leabra.Time) {
ly.Layer.AvgMaxAct(ltime)
ly.LeabraLay.(DeepLayer).AvgMaxActNoAttn(ltime)
}
// AvgMaxActNoAttn computes the average and max ActNoAttn stats
func (ly *Layer) AvgMaxActNoAttn(ltime *leabra.Time) {
for pi := range ly.DeepPools {
pl := &ly.Pools[pi]
dpl := &ly.DeepPools[pi]
dpl.ActNoAttn.Init()
for ni := pl.StIdx; ni < pl.EdIdx; ni++ {
dnr := &ly.DeepNeurs[ni]
dpl.ActNoAttn.UpdateVal(dnr.ActNoAttn, ni)
}
dpl.ActNoAttn.CalcAvg()
}
}
//////////////////////////////////////////////////////////////////////////////////////
// DeepBurst -- computed every cycle at end of standard Cycle in DeepBurst quarter
// DeepBurstFmAct updates DeepBurst layer 5 IB bursting value from current Act (superficial activation)
// Subject to thresholding.
func (ly *Layer) DeepBurstFmAct(ltime *leabra.Time) {
if !ly.DeepBurst.On || !ly.DeepBurst.IsBurstQtr(ltime.Quarter) {
return
}
lpl := &ly.DeepPools[0]
actMax := lpl.ActNoAttn.Max
actAvg := lpl.ActNoAttn.Avg
thr := actAvg + ly.DeepBurst.ThrRel*(actMax-actAvg)
thr = math32.Max(thr, ly.DeepBurst.ThrAbs)
for ni := range ly.DeepNeurs {
nrn := &ly.Neurons[ni]
if nrn.IsOff() {
continue
}
dnr := &ly.DeepNeurs[ni]
burst := float32(0)
if dnr.ActNoAttn > thr {
burst = dnr.ActNoAttn
}
dnr.DeepBurst = burst
}
}
// SendTRCBurstGeDelta sends change in DeepBurst activation since last sent, over BurstTRC
// projections.
func (ly *Layer) SendTRCBurstGeDelta(ltime *leabra.Time) {
if !ly.DeepBurst.On || !ly.DeepBurst.IsBurstQtr(ltime.Quarter) {
return
}
for ni := range ly.DeepNeurs {
nrn := &ly.Neurons[ni]
if nrn.IsOff() {
continue
}
dnr := &ly.DeepNeurs[ni]
if dnr.DeepBurst > ly.Act.OptThresh.Send {
delta := dnr.DeepBurst - dnr.DeepBurstSent
if math32.Abs(delta) > ly.Act.OptThresh.Delta {
for _, sp := range ly.SndPrjns {
if sp.IsOff() {
continue
}
pj := sp.(DeepPrjn)
ptyp := pj.Type()
if ptyp != BurstTRC {
continue
}
pj.SendTRCBurstGeDelta(ni, delta)
}
dnr.DeepBurstSent = dnr.DeepBurst
}
} else if dnr.DeepBurstSent > ly.Act.OptThresh.Send {
delta := -dnr.DeepBurstSent // un-send the last above-threshold activation to get back to 0
for _, sp := range ly.SndPrjns {
if sp.IsOff() {
continue
}
pj := sp.(DeepPrjn)
ptyp := pj.Type()
if ptyp != BurstTRC {
continue
}
pj.SendTRCBurstGeDelta(ni, delta)
}
dnr.DeepBurstSent = 0
}
}
}
// TRCBurstGeFmInc computes the TRCBurstGe input from sent values
func (ly *Layer) TRCBurstGeFmInc(ltime *leabra.Time) {
if !ly.DeepBurst.On || !ly.DeepBurst.IsBurstQtr(ltime.Quarter) {
return
}
for _, p := range ly.RcvPrjns {
if p.IsOff() {
continue
}
pj := p.(DeepPrjn)
ptyp := pj.Type()
if ptyp != BurstTRC {
continue
}
pj.RecvTRCBurstGeInc()
}
// note: full integration of Inc happens next cycle..
}
// AvgMaxTRCBurstGe computes the average and max TRCBurstGe stats
func (ly *Layer) AvgMaxTRCBurstGe(ltime *leabra.Time) {
for pi := range ly.DeepPools {
pl := &ly.Pools[pi]
dpl := &ly.DeepPools[pi]
dpl.TRCBurstGe.Init()
for ni := pl.StIdx; ni < pl.EdIdx; ni++ {
dnr := &ly.DeepNeurs[ni]
dpl.TRCBurstGe.UpdateVal(dnr.TRCBurstGe, ni)
}
dpl.TRCBurstGe.CalcAvg()
}
}
//////////////////////////////////////////////////////////////////////////////////////
// DeepCtxt -- once after DeepBurst quarter
// SendDeepCtxtGe sends full DeepBurst activation over BurstCtxt projections to integrate
// DeepCtxtGe excitatory conductance on deep layers.
// This must be called at the end of the DeepBurst quarter for this layer.
func (ly *Layer) SendDeepCtxtGe(ltime *leabra.Time) {
if !ly.DeepBurst.On || !ly.DeepBurst.IsBurstQtr(ltime.Quarter) {
return
}
for ni := range ly.DeepNeurs {
nrn := &ly.Neurons[ni]
if nrn.IsOff() {
continue
}
dnr := &ly.DeepNeurs[ni]
if dnr.DeepBurst > ly.Act.OptThresh.Send {
for _, sp := range ly.SndPrjns {
if sp.IsOff() {
continue
}
pj := sp.(DeepPrjn)
ptyp := pj.Type()
if ptyp != BurstCtxt {
continue
}
pj.SendDeepCtxtGe(ni, dnr.DeepBurst)
}
}
}
}
// DeepCtxtFmGe integrates new DeepCtxtGe excitatory conductance from projections, and computes
// overall DeepCtxt value, only on Deep layers.
// This must be called at the end of the DeepBurst quarter for this layer, after SendDeepCtxtGe.
func (ly *Layer) DeepCtxtFmGe(ltime *leabra.Time) {
if ly.Typ != Deep || !ly.DeepBurst.IsBurstQtr(ltime.Quarter) {
return
}
for ni := range ly.DeepNeurs {
dnr := &ly.DeepNeurs[ni]
dnr.DeepCtxtGe = 0
}
for _, p := range ly.RcvPrjns {
if p.IsOff() {
continue
}
pj := p.(DeepPrjn)
ptyp := pj.Type()
if ptyp != BurstCtxt {
continue
}
pj.RecvDeepCtxtGeInc()
}
for ni := range ly.DeepNeurs {
nrn := &ly.Neurons[ni]
if nrn.IsOff() {
continue
}
dnr := &ly.DeepNeurs[ni]
dnr.DeepCtxt = ly.DeepCtxt.DeepCtxtFmGe(dnr.DeepCtxtGe, dnr.DeepCtxt)
}
}
// QuarterFinal does updating after end of a quarter
func (ly *Layer) QuarterFinal(ltime *leabra.Time) {
ly.Layer.QuarterFinal(ltime)
ly.LeabraLay.(DeepLayer).DeepBurstPrv(ltime)
}
// DeepBurstPrv saves DeepBurst as DeepBurstPrv
func (ly *Layer) DeepBurstPrv(ltime *leabra.Time) {
if !ly.DeepBurst.On || !ly.DeepBurst.NextIsBurstQtr(ltime.Quarter) {
return
}
for ni := range ly.DeepNeurs {
dnr := &ly.DeepNeurs[ni]
dnr.DeepBurstPrv = dnr.DeepBurst
}
}
//////////////////////////////////////////////////////////////////////////////////////
// LayerType
// DeepLeabra extensions to the emer.LayerType types
//go:generate stringer -type=LayerType
var KiT_LayerType = kit.Enums.AddEnum(LayerTypeN, false, nil)
// The DeepLeabra layer types
const (
// Super are superficial-layer neurons, which also compute DeepBurst activation as a
// thresholded version of superficial activation, and send that to both TRC (for plus
// phase outcome) and Deep layers (for DeepCtxt temporal context).
Super emer.LayerType = emer.LayerTypeN + iota
// Deep are deep-layer neurons, reflecting activation of layer 6 regular spiking
// CT corticothalamic neurons, which drive both attention in Super (via DeepAttn
// projections) and predictions in TRC (Pulvinar) via standard projections.
Deep
// TRC are thalamic relay cell neurons, typically in the Pulvinar, which alternately reflect
// predictions driven by Deep layer projections, and actual outcomes driven by BurstTRC
// projections from corresponding Super layer neurons that provide strong driving inputs to
// TRC neurons.
TRC
LayerTypeN
)