vspatch learning still not good

[rcoreilly]

https://github.com/emer/axon/discussions/306#discussioncomment-6918314

[rcoreilly]

Finally , vspatch seems to be doing well, using key ideas from #306:

• GiBase initialized per neuron so there is a spectrum of excitability -- more inhibited get recruited later.
• Sigmoid derivative on RLRate prevents weights from saturating
• fixed major bug in adapting the VSPatch threshold -- was using raw VSPatch activity vs. thresholded.

In boa:

{Sel: ".VSPatchLayer", Desc: "",
    Params: params.Params{
        "Layer.Learn.NeuroMod.DipGain":     "1",    // if < 1, overshoots, more -DA; but quick extinct in pvlv
        "Layer.Learn.RLRate.SigmoidMin":    "0.01", // 0.05 def
        "Layer.VSPatch.Gain":               "3", // 3 smoother than higher
        "Layer.VSPatch.ThrInit":            "0.2", // could be a bit lower
        "Layer.VSPatch.ThrLRate":           "0.002",
        "Layer.VSPatch.ThrNonRew":          "10", // 10 works to prevent RewPred_NR
        "Layer.Learn.TrgAvgAct.GiBaseInit": "0.5", // 0.5 gives reasonably sparse pattern -- seems good
    }},

[rcoreilly]

This is all working much better now: see https://github.com/emer/axon/pull/302 for updates.