-
Notifications
You must be signed in to change notification settings - Fork 8
/
params.go
219 lines (216 loc) · 7.62 KB
/
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
// Copyright (c) 2022, 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/v2/netparams"
"github.com/emer/emergent/v2/params"
)
// ParamSets is the active set of parameters -- Base is always applied,
// and others can be optionally selected to apply on top of that
var ParamSets = netparams.Sets{
"Base": {
{Sel: "Layer", Desc: "generic params for all layers",
Params: params.Params{
"Layer.Acts.Clamp.Ge": "1.5",
}},
{Sel: ".PFCLayer", Desc: "pfc layers: slower trgavgact",
Params: params.Params{
"Layer.Learn.TrgAvgAct.SynScaleRate": "0.0002", // also now set by default
}},
{Sel: ".PTMaintLayer", Desc: "time integration params",
Params: params.Params{
// "Layer.Inhib.Layer.Gi": "2.4",
// "Layer.Inhib.Pool.Gi": "2.4",
"Layer.Acts.Dend.ModGain": "1.5", // 2 min -- reduces maint early
"Layer.Learn.NeuroMod.AChDisInhib": "0.0", // not much effect here..
}},
{Sel: ".BLALayer", Desc: "",
Params: params.Params{
"Layer.Learn.NeuroMod.DAModGain": "0.5",
}},
{Sel: ".PTPredLayer", Desc: "",
Params: params.Params{
"Layer.Inhib.ActAvg.Nominal": "0.1",
"Layer.CT.GeGain": "0.05", // 0.05 key for stronger activity
"Layer.CT.DecayTau": "50",
"Layer.Learn.NeuroMod.AChDisInhib": "0", // 0.2, 0.5 not much diff
}},
{Sel: ".CS", Desc: "need to adjust Nominal for number of CSs -- now down automatically",
Params: params.Params{
"Layer.Inhib.ActAvg.Nominal": "0.1", // 0.1 for 4, divide by N/4 from there
}},
// {Sel: "#OFCpos", Desc: "",
// Params: params.Params{
// "Layer.Inhib.Pool.Gi": "1",
// }},
// {Sel: "#OFCposPT", Desc: "",
// Params: params.Params{
// "Layer.Inhib.Pool.Gi": "0.5",
// }},
{Sel: "#OFCposPTp", Desc: "",
Params: params.Params{
"Layer.Inhib.ActAvg.Nominal": "0.1", // 0.1 -- affects how strongly BLA is driven -- key param
"Layer.Inhib.Pool.Gi": "1.4",
}},
{Sel: "#ILposPTp", Desc: "",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.2",
}},
{Sel: "#ILnegPTp", Desc: "",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.2",
}},
{Sel: "#OFCneg", Desc: "",
Params: params.Params{
"Layer.Inhib.ActAvg.Nominal": "0.1",
// "Layer.Inhib.Layer.Gi": "0.5", // weaker in general so needs to be lower
}},
// {Sel: "#OFCnegPT", Desc: "",
// Params: params.Params{
// "Layer.Inhib.ActAvg.Nominal": "0.2",
// "Layer.Inhib.Pool.Gi": "3.0",
// }},
// {Sel: "#OFCnegPTp", Desc: "",
// Params: params.Params{
// "Layer.Inhib.Pool.Gi": "1.4",
// }},
// {Sel: "#ILpos", Desc: "",
// Params: params.Params{
// "Layer.Inhib.Pool.Gi": "1",
// }},
{Sel: ".VSMatrixLayer", Desc: "vs mtx",
Params: params.Params{
"Layer.Inhib.Layer.On": "false", // todo: explore -- could be bad for gating
"Layer.Inhib.Pool.Gi": "0.5", // go lower, get more inhib from elsewhere?
"Layer.Inhib.Pool.FB": "0",
"Layer.Acts.Dend.ModGain": "1", // todo: 2 is default
"Layer.Acts.Kir.Gbar": "2",
"Layer.Learn.NeuroMod.BurstGain": "1",
"Layer.Learn.NeuroMod.DAModGain": "0", // no bias is better!
"Layer.Learn.RLRate.SigmoidMin": "0.001", // 0.01 better than .05
}},
{Sel: "#BLAposAcqD1", Desc: "",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "2.4", // 2.2 not enough to knock out novelty
"Layer.Inhib.Pool.Gi": "1",
}},
{Sel: "#BLAnegAcqD2", Desc: "",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.2", // weaker
}},
{Sel: ".VSPatchLayer", Desc: "",
Params: params.Params{
"Layer.Inhib.Pool.Gi": "0.5", // 0.5 ok?
"Layer.Inhib.Pool.FB": "0", // only fb
"Layer.Learn.NeuroMod.DipGain": "1", // if < 1, overshoots, more -DA
"Layer.Learn.NeuroMod.BurstGain": "1",
"Layer.Learn.RLRate.SigmoidMin": "0.01", // 0.01 > 0.05 def
"Layer.Learn.TrgAvgAct.GiBaseInit": "0", // 0.2 gets too diffuse
}},
{Sel: ".LDTLayer", Desc: "",
Params: params.Params{
"Layer.LDT.MaintInhib": "2.0", // 0.95 is too weak -- depends on activity..
}},
{Sel: "#SC", Desc: "",
Params: params.Params{
"Layer.Acts.KNa.Slow.Max": "0.8", // .8 reliable decreases -- could go higher
}},
////////////////////////////////////////////
// Cortical Prjns
{Sel: ".PFCPrjn", Desc: "pfc prjn params -- more robust to long-term training",
Params: params.Params{
"Prjn.Learn.Trace.SubMean": "1", // 1 > 0 for long-term stability
"Prjn.Learn.LRate.Base": "0.01", // 0.04 def; 0.02 more stable; 0.01 even more
}},
{Sel: ".PTtoPred", Desc: "stronger drive on pt pred",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1",
}},
{Sel: "#BLAposAcqD1ToOFCpos", Desc: "stronger",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1.5", // stronger = bad later
}},
{Sel: "#OFCposToILpos", Desc: "stronger",
Params: params.Params{
"Prjn.PrjnScale.Abs": "3",
}},
{Sel: ".USToBLAExtInhib", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "2",
}},
{Sel: "#ILposToPLutil", Desc: "not good to make this stronger",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1", // todo: try 3?
}},
{Sel: ".MToACC", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "3",
}},
// {Sel: ".PTSelfMaint", Desc: "",
// Params: params.Params{
// "Prjn.PrjnScale.Abs": "4",
// "Prjn.Learn.LRate.Base": "0.0001", // this is not a problem
// }},
////////////////////////////////////////////
// Rubicon Prjns
{Sel: ".VSMatrixPrjn", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1.5", // 3 orig
"Prjn.Learn.Trace.LearnThr": "0.1",
"Prjn.Learn.LRate.Base": "0.01", // 0.05 def
}},
{Sel: ".ToSC", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "2",
}},
{Sel: ".DrivesToMtx", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "1",
}},
{Sel: ".BLAExtPrjn", Desc: "ext learns relatively fast",
Params: params.Params{
"Prjn.Learn.LRate.Base": "0.005",
}},
{Sel: ".BLAAcqToGo", Desc: "must dominate",
Params: params.Params{
"Prjn.PrjnScale.Rel": "1",
"Prjn.PrjnScale.Abs": "4",
}},
{Sel: ".BLAExtToAcq", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "0.5", // note: key param -- 0.5 > 1
}},
{Sel: ".PFCToVSMtx", Desc: "contextual, should be weaker",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1", // 0.1 def
}},
{Sel: ".VSPatchPrjn", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "3", // 4 > 3 > 2 -- key for rapid learning
"Prjn.Learn.Trace.LearnThr": "0",
"Prjn.Learn.LRate.Base": "0.02", // 0.02 needed in test
}},
{Sel: "#CSToBLAposAcqD1", Desc: "",
Params: params.Params{
"Prjn.Learn.LRate.Base": "0.1", // was 0.5 -- too fast!?
}},
{Sel: ".SuperToThal", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "4", // 4 def
}},
{Sel: ".SuperToPT", Desc: "",
Params: params.Params{
"Prjn.PrjnScale.Abs": "0.5", // 0.5 def
}},
{Sel: ".GPiToBGThal", Desc: "inhibition from GPi to MD",
Params: params.Params{
"Prjn.PrjnScale.Abs": "5", // with new mod, this can be stronger
}},
{Sel: ".BLAFromNovel", Desc: "Note: this setting is overwritten in boa.go ApplyParams",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1", // weak rel to not dilute rest of bla prjns
"Prjn.PrjnScale.Abs": "3", // 2 is good for .CS nominal .1, but 3 needed for .03
}},
},
}