-
Notifications
You must be signed in to change notification settings - Fork 21
/
def_params.go
296 lines (293 loc) · 10.6 KB
/
def_params.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
// Copyright (c) 2020, 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 main
import "github.com/emer/emergent/params"
// ParamSets is the default set of parameters -- Base is always applied, and others can be optionally
// selected to apply on top of that
var ParamSets = params.Sets{
{Name: "Base", Desc: "these are the best params", Sheets: params.Sheets{
"Network": ¶ms.Sheet{
{Sel: "Prjn", Desc: "keeping default params for generic prjns",
Params: params.Params{
"Prjn.Learn.Momentum.On": "true",
"Prjn.Learn.Norm.On": "true",
"Prjn.Learn.WtBal.On": "false",
}},
{Sel: ".EcCa1Prjn", Desc: "encoder projections -- no norm, moment",
Params: params.Params{
"Prjn.Learn.Lrate": "0.04",
"Prjn.Learn.Momentum.On": "false",
"Prjn.Learn.Norm.On": "false",
"Prjn.Learn.WtBal.On": "true", // counteracting hogging
//"Prjn.Learn.XCal.SetLLrn": "true", // bcm now avail, comment out = default LLrn
//"Prjn.Learn.XCal.LLrn": "0", // 0 = turn off BCM, must with SetLLrn = true
}},
{Sel: ".HippoCHL", Desc: "hippo CHL projections -- no norm, moment, but YES wtbal = sig better",
Params: params.Params{
"Prjn.CHL.Hebb": "0.01", // .01 > .05? > .1?
"Prjn.Learn.Lrate": "0.2", // .2 probably better? .4 was prev default
"Prjn.Learn.Momentum.On": "false",
"Prjn.Learn.Norm.On": "false",
"Prjn.Learn.WtBal.On": "true",
}},
{Sel: ".PPath", Desc: "performant path, new Dg error-driven EcCa1Prjn prjns",
Params: params.Params{
"Prjn.Learn.Lrate": "0.15", // err driven: .15 > .2 > .25 > .1
"Prjn.Learn.Momentum.On": "false",
"Prjn.Learn.Norm.On": "false",
"Prjn.Learn.WtBal.On": "true",
//"Prjn.Learn.XCal.SetLLrn": "true", // bcm now avail, comment out = default LLrn
//"Prjn.Learn.XCal.LLrn": "0", // 0 = turn off BCM, must with SetLLrn = true
}},
{Sel: "#CA1ToECout", Desc: "extra strong from CA1 to ECout",
Params: params.Params{
"Prjn.WtScale.Abs": "4.0", // 4 > 6 > 2 (fails)
}},
{Sel: "#InputToECin", Desc: "one-to-one input to EC",
Params: params.Params{
"Prjn.Learn.Learn": "false",
"Prjn.WtInit.Mean": "0.8",
"Prjn.WtInit.Var": "0.0",
}},
{Sel: "#ECoutToECin", Desc: "one-to-one out to in",
Params: params.Params{
"Prjn.Learn.Learn": "false",
"Prjn.WtInit.Mean": "0.9",
"Prjn.WtInit.Var": "0.01",
"Prjn.WtScale.Rel": "0.5", // .5 = .3? > .8 (fails); zycyc test this
}},
{Sel: "#DGToCA3", Desc: "Mossy fibers: strong, non-learning",
Params: params.Params{
"Prjn.Learn.Learn": "false", // learning here definitely does NOT work!
"Prjn.WtInit.Mean": "0.9",
"Prjn.WtInit.Var": "0.01",
"Prjn.WtScale.Rel": "4", // err del 4: 4 > 6 > 8
//"Prjn.WtScale.Abs": "1.5", // zycyc, test if abs activation was not enough
}},
//{Sel: "#ECinToCA3", Desc: "ECin Perforant Path",
// Params: params.Params{
// "Prjn.WtScale.Abs": "1.5", // zycyc, test if abs activation was not enough
// }},
{Sel: "#CA3ToCA3", Desc: "CA3 recurrent cons: rel=2 still the best",
Params: params.Params{
"Prjn.WtScale.Rel": "2", // 2 > 1 > .5 = .1
"Prjn.Learn.Lrate": "0.1", // .1 > .08 (close) > .15 > .2 > .04;
//"Prjn.WtScale.Abs": "1.5", // zycyc, test if abs activation was not enough
}},
{Sel: "#ECinToDG", Desc: "DG learning is surprisingly critical: maxed out fast, hebbian works best",
Params: params.Params{
"Prjn.Learn.Learn": "true", // absolutely essential to have on! learning slow if off.
"Prjn.CHL.Hebb": "0.2", // .2 seems good
"Prjn.CHL.SAvgCor": "0.1", // 0.01 = 0.05 = .1 > .2 > .3 > .4 (listlize 20-100)
"Prjn.CHL.MinusQ1": "true", // dg self err slightly better
"Prjn.Learn.Lrate": "0.05", // .05 > .1 > .2 > .4; .01 less interference more learning time - key tradeoff param, .05 best for list20-100
"Prjn.Learn.Momentum.On": "false",
"Prjn.Learn.Norm.On": "false",
"Prjn.Learn.WtBal.On": "true",
}},
{Sel: "#CA3ToCA1", Desc: "Schaffer collaterals -- slower, less hebb",
Params: params.Params{
"Prjn.CHL.Hebb": "0.01", // .01 > .005 > .02 > .002 > .001 > .05 (crazy)
"Prjn.CHL.SAvgCor": "0.4",
"Prjn.Learn.Lrate": "0.1", // CHL: .1 =~ .08 > .15 > .2, .05 (sig worse)
"Prjn.Learn.Momentum.On": "false",
"Prjn.Learn.Norm.On": "false",
"Prjn.Learn.WtBal.On": "true",
//"Prjn.WtScale.Abs": "1.5", // zycyc, test if abs activation was not enough
}},
//{Sel: "#ECinToCA1", Desc: "ECin Perforant Path",
// Params: params.Params{
// "Prjn.WtScale.Abs": "1.5", // zycyc, test if abs activation was not enough
// }},
{Sel: "#ECoutToCA1", Desc: "ECout Perforant Path",
Params: params.Params{
"Prjn.WtScale.Rel": "0.3", // Back proj should generally be very weak but we're specifically setting this here bc others are set already
}},
{Sel: ".EC", Desc: "all EC layers: only pools, no layer-level -- now for EC3 and EC5",
Params: params.Params{
"Layer.Act.Gbar.L": "0.1",
"Layer.Inhib.ActAvg.Init": "0.2",
"Layer.Inhib.Layer.On": "false",
"Layer.Inhib.Pool.Gi": "2.0",
"Layer.Inhib.Pool.On": "true",
}},
{Sel: "#DG", Desc: "very sparse = high inhibition",
Params: params.Params{
"Layer.Inhib.ActAvg.Init": "0.01",
"Layer.Inhib.Layer.Gi": "3.8", // 3.8 > 3.6 > 4.0 (too far -- tanks)
}},
{Sel: "#CA3", Desc: "sparse = high inhibition",
Params: params.Params{
"Layer.Inhib.ActAvg.Init": "0.02",
"Layer.Inhib.Layer.Gi": "2.8", // 2.8 = 3.0 really -- some better, some worse
"Layer.Learn.AvgL.Gain": "2.5", // stick with 2.5
}},
{Sel: "#CA1", Desc: "CA1 only Pools",
Params: params.Params{
"Layer.Inhib.ActAvg.Init": "0.1",
"Layer.Inhib.Layer.On": "false",
"Layer.Inhib.Pool.On": "true",
"Layer.Inhib.Pool.Gi": "2.4", // 2.4 > 2.2 > 2.6 > 2.8 -- 2.4 better *for small net* but not for larger!
"Layer.Learn.AvgL.Gain": "2.5", // 2.5 > 2 > 3
//"Layer.Inhib.ActAvg.UseFirst": "false", // first activity is too low, throws off scaling, from Randy, zycyc: do we need this?
}},
},
// NOTE: it is essential not to put Pat / Hip params here, as we have to use Base
// to initialize the network every time, even if it is a different size..
}},
{Name: "List010", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "10",
}},
},
}},
{Name: "List020", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "20",
}},
},
}},
{Name: "List030", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "30",
}},
},
}},
{Name: "List040", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "40",
}},
},
}},
{Name: "List050", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "50",
}},
},
}},
{Name: "List060", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "60",
}},
},
}},
{Name: "List070", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "70",
}},
},
}},
{Name: "List080", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "80",
}},
},
}},
{Name: "List090", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "90",
}},
},
}},
{Name: "List100", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "100",
}},
},
}},
{Name: "List125", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "125",
}},
},
}},
{Name: "List150", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "150",
}},
},
}},
{Name: "List175", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "175",
}},
},
}},
{Name: "List200", Desc: "list size", Sheets: params.Sheets{
"Pat": ¶ms.Sheet{
{Sel: "PatParams", Desc: "pattern params",
Params: params.Params{
"PatParams.ListSize": "200",
}},
},
}},
{Name: "SmallHip", Desc: "hippo size", Sheets: params.Sheets{
"Hip": ¶ms.Sheet{
{Sel: "HipParams", Desc: "hip sizes",
Params: params.Params{
"HipParams.ECPool.Y": "7",
"HipParams.ECPool.X": "7",
"HipParams.CA1Pool.Y": "10",
"HipParams.CA1Pool.X": "10",
"HipParams.CA3Size.Y": "20",
"HipParams.CA3Size.X": "20",
"HipParams.DGRatio": "2.236", // 1.5 before, sqrt(5) aligns with Ketz et al. 2013
}},
},
}},
{Name: "MedHip", Desc: "hippo size", Sheets: params.Sheets{
"Hip": ¶ms.Sheet{
{Sel: "HipParams", Desc: "hip sizes",
Params: params.Params{
"HipParams.ECPool.Y": "7",
"HipParams.ECPool.X": "7",
"HipParams.CA1Pool.Y": "15",
"HipParams.CA1Pool.X": "15",
"HipParams.CA3Size.Y": "30",
"HipParams.CA3Size.X": "30",
"HipParams.DGRatio": "2.236", // 1.5 before
}},
},
}},
{Name: "BigHip", Desc: "hippo size", Sheets: params.Sheets{
"Hip": ¶ms.Sheet{
{Sel: "HipParams", Desc: "hip sizes",
Params: params.Params{
"HipParams.ECPool.Y": "7",
"HipParams.ECPool.X": "7",
"HipParams.CA1Pool.Y": "20",
"HipParams.CA1Pool.X": "20",
"HipParams.CA3Size.Y": "40",
"HipParams.CA3Size.X": "40",
"HipParams.DGRatio": "2.236", // 1.5 before
}},
},
}},
}