-
Notifications
You must be signed in to change notification settings - Fork 7
/
params.go
262 lines (258 loc) · 9.28 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
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
// 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 default 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: "clamp gain makes big diff on overall excitation, gating propensity",
Params: params.Params{
"Layer.Acts.Clamp.Ge": "1.0", // 1.5 is def, was 0.6 (too low)
}},
{Sel: ".VBG", Desc: "all ModACh",
Params: params.Params{
"Layer.Acts.Dend.ModACh": "true",
}},
{Sel: ".MatrixLayer", Desc: "all mtx",
Params: params.Params{
"Layer.Inhib.Pool.On": "false",
"Layer.Inhib.Layer.Gi": "0.5", // 0.5 needed for differentiated reps
"Layer.Inhib.ActAvg.Nominal": "0.25",
"Layer.Matrix.GateThr": "0.05", // todo: .01 should be new default
"Layer.Matrix.IsVS": "true", // key for resetting urgency
"Layer.Learn.RLRate.On": "true", // only used for non-rew trials -- key
"Layer.Learn.RLRate.Diff": "false",
"Layer.Learn.RLRate.SigmoidMin": "0.01", // 0.01 better than .05
"Layer.Learn.TrgAvgAct.RescaleOn": "true",
"Layer.Learn.NeuroMod.BurstGain": "0.1", // 1 def -- must be smaller given rew dynamics
"Layer.Learn.NeuroMod.DAModGain": "0", // strongly biases the gating
}},
{Sel: ".VSTNLayer", Desc: "all VSTN",
Params: params.Params{
"Layer.Acts.Init.GeBase": "0.1", // todo: re-param with more stn, increase..
"Layer.Acts.SKCa.CaRDecayTau": "80",
}},
{Sel: ".PTMaintLayer", Desc: "time integration params",
Params: params.Params{
// "Layer.Inhib.Layer.Gi": "3.2", // 3.2 def
"Layer.Acts.Dend.ModGain": "1.5", // 1.5 def
"Layer.Acts.Kir.Gbar": "0", // no real diff here over range 0-10
"Layer.Acts.Dend.ModACh": "true",
}},
{Sel: ".ACC", Desc: "manipulate noise to see about integration over time",
Params: params.Params{
"Layer.Acts.Noise.On": "false",
"Layer.Acts.Noise.Ge": "0.1", // 0.1 is visibly impactful
"Layer.Acts.Noise.Gi": "0.01", // 0.01 -- if to strong, rep becomes very weak
}},
////////////////////////////////////////////
// Paths
{Sel: ".VSMatrixPath", Desc: "",
Params: params.Params{
"Path.Learn.LRate.Base": "0.01", // 0.01, vs .02 default
"Path.Learn.Trace.LearnThr": "0.1", // prevents learning below this thr: preserves low act
"Path.Matrix.VSRewLearn": "false", // significantly cleaner
"Path.SWts.Adapt.On": "false", // not much diff: false >= true
},
Hypers: params.Hypers{
"Path.Learn.LRate.Base": {"Tweak": "-"},
"Path.Learn.Trace.LearnThr": {"Tweak": "-"},
}},
{Sel: "#UrgencyToVMtxGo", Desc: "strong urgency factor",
Params: params.Params{
"Path.PathScale.Rel": "0.1", // don't dilute from others
"Path.PathScale.Abs": "0", // todo: is misbehaving here
"Path.Learn.Learn": "false",
}},
{Sel: ".SuperToPT", Desc: "one-to-one from super",
Params: params.Params{
"Path.PathScale.Abs": "0.5",
}},
{Sel: ".SuperToThal", Desc: "",
Params: params.Params{
"Path.PathScale.Abs": "3.0", // was 4
}},
{Sel: ".ACCToVMtx", Desc: "",
Params: params.Params{
"Path.PathScale.Abs": "1.5", // 1.5 good; 1.8 causes some breakthrough
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VMtxNoToVMtxGo", Desc: "",
Params: params.Params{
"Path.PathScale.Rel": "0.05",
"Path.PathScale.Abs": "1",
"Path.Learn.Learn": "false",
},
Hypers: params.Hypers{
"Path.PathScale.Rel": {"Tweak": "log"},
}},
{Sel: "#VGPi", Desc: "",
Params: params.Params{
"Layer.Acts.Init.GeBase": "0.3", // 0.3 > 0.2, 0.1
"Layer.Acts.Init.GeVar": "0.1",
},
Hypers: params.Hypers{
"Layer.Acts.Init.GeBase": {"Tweak": "-"},
}},
{Sel: "#VSTNToVGPi", Desc: "strong initial phasic activation",
Params: params.Params{
"Path.PathScale.Abs": ".2",
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
},
}
// ParamSetsDefs contains the full set of parameters, many of which are at default values
// and have informed the default values in the first place.
var ParamSetsDefs = netparams.Sets{
"Defaults": {
{Sel: ".MatrixLayer", Desc: "all mtx",
Params: params.Params{
"Layer.Inhib.Pool.On": "false",
"Layer.Inhib.Layer.Gi": "0.5",
"Layer.Inhib.Layer.FB": "0",
"Layer.Matrix.GateThr": "0.05", // .05 default
"Layer.Acts.Kir.Gbar": "10", // 10 > 5 > 20
"Layer.Acts.GabaB.Gbar": "0",
"Layer.Acts.NMDA.Gbar": "0.006", // 0.006 default, necessary (0 very bad)
"Layer.Learn.NeuroMod.AChLRateMod": "0", // no diff here -- always ACh
"Layer.Learn.NeuroMod.BurstGain": "0.1", // 0.1 == 0.2 > 0.05 > 0.5; only for weird rew case here; 1 def
},
Hypers: params.Hypers{
"Layer.Learn.NeuroMod.BurstGain": {"Tweak": "-"},
"Layer.Acts.Kir.Gbar": {"Tweak": "-"},
"Layer.Acts.NMDA.Gbar": {"Tweak": "-"},
"Layer.Inhib.Layer.Gi": {"Tweak": "-"},
}},
{Sel: ".VSTNLayer", Desc: "all VSTN",
Params: params.Params{
"Layer.Acts.Init.GeBase": "0.1",
"Layer.Acts.Kir.Gbar": "10", // 10 > 5 > 2 -- key for pause
"Layer.Acts.SKCa.Gbar": "2", // 2 > 5 >> 1 (for Kir = 10)
"Layer.Inhib.Layer.On": "true", // really no inhib neurons here. all VGPePr
"Layer.Learn.NeuroMod.AChDisInhib": "0",
},
Hypers: params.Hypers{
"Layer.Acts.Init.GeBase": {"Tweak": "-"},
"Layer.Acts.Kir.Gbar": {"Tweak": "-"},
"Layer.Acts.SKCa.Gbar": {"Tweak": "-"},
}},
{Sel: "#VGPePr", Desc: "prototypical",
Params: params.Params{
"Layer.Acts.Init.GeBase": "0.4", // 0.4 > 0.3, 0.5
"Layer.Acts.Init.GeVar": "0.2",
},
Hypers: params.Hypers{
"Layer.Acts.Init.GeBase": {"Tweak": "-"},
}},
{Sel: "#VGPeAk", Desc: "arkypallidal",
Params: params.Params{
"Layer.Acts.Init.GeBase": "0.2", // 0.2 > 0.3, 0.1
"Layer.Acts.Init.GeVar": "0.1", // 0.1 == 0.2 > 0.05
},
Hypers: params.Hypers{
"Layer.Acts.Init.GeBase": {"Tweak": "-"},
"Layer.Acts.Init.GeVar": {"Tweak": "-"},
}},
{Sel: "#VGPi", Desc: "",
Params: params.Params{
"Layer.Acts.Init.GeBase": "0.3", // 0.3 > 0.2, 0.1
"Layer.Acts.Init.GeVar": "0.1",
},
Hypers: params.Hypers{
"Layer.Acts.Init.GeBase": {"Tweak": "-"},
}},
{Sel: ".VGPeAkToVMtx", Desc: "go disinhibition",
Params: params.Params{
"Path.PathScale.Abs": "3", // 3 >= 2, 4
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VMtxGoToVGPeAk", Desc: "go inhibition",
Params: params.Params{
"Path.PathScale.Abs": ".5", // stronger = more binary
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VGPePrToVSTN", Desc: "enough to kick off the ping-pong dynamics for VSTN.",
Params: params.Params{
"Path.PathScale.Abs": "0.5",
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VSTNToVGPePr", Desc: "stronger VSTN -> VGPePr to kick it high at start",
Params: params.Params{
"Path.PathScale.Abs": "0.5",
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VSTNToVGPeAk", Desc: "this is weak biologically -- could try 0",
Params: params.Params{
"Path.PathScale.Abs": "0.1",
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VMtxNoToVGPePr", Desc: "proto = primary classical NoGo pathway",
Params: params.Params{
"Path.PathScale.Abs": "1", // 1 fully inhibits Pr
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VGPePrToVGPePr", Desc: "self-inhib -- only source of self reg",
Params: params.Params{
"Path.PathScale.Abs": "4", // 4 best for DS
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VGPePrToVGPeAk", Desc: "just enough to knock down in baseline state",
Params: params.Params{
"Path.PathScale.Abs": "1",
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VMtxGoToVGPi", Desc: "go influence on gating",
Params: params.Params{
"Path.PathScale.Abs": ".2", // .1 too weak
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VGPePrToVGPi", Desc: "nogo influence on gating -- decreasing produces more graded function of Go",
Params: params.Params{
"Path.PathScale.Abs": "1", // 2 is much worse.. keep at 1
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VSTNToVGPi", Desc: "strong initial phasic activation",
Params: params.Params{
"Path.PathScale.Abs": ".2",
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
{Sel: "#VGPiToACCPosVM", Desc: "final inhibition",
Params: params.Params{
"Path.PathScale.Abs": "5", // needs to be very strong -- 5
},
Hypers: params.Hypers{
"Path.PathScale.Abs": {"Tweak": "-"},
}},
},
}