first impl of linear regression synaptic ca gives a 4x speedup and learning performance is consistently _improved_ relative to prior, across ra25, deep_fsa, and objrec, which usually means that it is genuinely better. Lots more work to be done to explore the space but this is an encouraging start!

Author: rcoreilly

axon/enumgen.go

@@ -682,9 +682,26 @@ func (ls *LRateParams) Init() {
 	ls.UpdateEff()
 }
 
+// SynCaFuns are different ways of computing synaptic calcium (experimental)
+type SynCaFuns int32 //enums:enum
+
+const (
+	// StdSynCa uses standard synaptic calcium integration method
+	StdSynCa SynCaFuns = iota
+
+	// LinearSynCa uses linear regression generated calcium integration (much faster)
+	LinearSynCa
+
+	// NeurSynCa uses simple product of separately-integrated neuron values (much faster)
+	NeurSynCa
+)
+
 // TraceParams manages parameters associated with temporal trace learning
 type TraceParams struct {
 
+	// how to compute the synaptic calcium (experimental)
+	SynCa SynCaFuns
+
 	// time constant for integrating trace over theta cycle timescales -- governs the decay rate of syanptic trace
 	Tau float32 `default:"1,2,4"`
 
@@ -696,9 +713,12 @@ type TraceParams struct {
 
 	// rate = 1 / tau
 	Dt float32 `view:"-" json:"-" xml:"-" edit:"-"`
+
+	pad, pad1, pad2 float32
 }
 
 func (tp *TraceParams) Defaults() {
+	tp.SynCa = LinearSynCa
 	tp.Tau = 1
 	tp.SubMean = 0
 	tp.LearnThr = 0

axon/learn.go

@@ -277,8 +277,7 @@ func (pj *PathParams) GatherSpikes(ctx *Context, ly *LayerParams, ni, di uint32,
 // DoSynCa returns false if should not do synaptic-level calcium updating.
 // Done by default in Cortex, not for some other special pathway types.
 func (pj *PathParams) DoSynCa() bool {
-	if pj.PathType == RWPath || pj.PathType == TDPredPath || pj.PathType == VSMatrixPath ||
-		pj.PathType == DSMatrixPath || pj.PathType == VSPatchPath || pj.PathType == BLAPath {
+	if pj.Learn.Trace.SynCa != StdSynCa || pj.PathType == RWPath || pj.PathType == TDPredPath || pj.PathType == VSMatrixPath || pj.PathType == DSMatrixPath || pj.PathType == VSPatchPath || pj.PathType == BLAPath || pj.Learn.Hebb.On.IsTrue() {
 		return false
 	}
 	return true
@@ -338,28 +337,36 @@ func (pj *PathParams) DWtSyn(ctx *Context, syni, si, ri, di uint32, layPool, sub
 // Uses synaptically integrated spiking, computed at the Theta cycle interval.
 // This is the trace version for hidden units, and uses syn CaP - CaD for targets.
 func (pj *PathParams) DWtSynCortex(ctx *Context, syni, si, ri, di uint32, layPool, subPool *Pool, isTarget bool) {
-	// credit assignment part
-	caUpT := SynCaV(ctx, syni, di, CaUpT)                                // time of last update
-	syCaM := SynCaV(ctx, syni, di, CaM)                                  // fast time scale
-	syCaP := SynCaV(ctx, syni, di, CaP)                                  // slower but still fast time scale, drives Potentiation
-	syCaD := SynCaV(ctx, syni, di, CaD)                                  // slow time scale, drives Depression (one trial = 200 cycles)
-	pj.Learn.KinaseCa.CurCa(ctx.SynCaCtr, caUpT, &syCaM, &syCaP, &syCaD) // always update, getting current Ca (just optimization)
-
-	rb0 := NrnV(ctx, ri, di, SpkBin0)
-	sb0 := NrnV(ctx, si, di, SpkBin0)
-	rb1 := NrnV(ctx, ri, di, SpkBin1)
-	sb1 := NrnV(ctx, si, di, SpkBin1)
-	rb2 := NrnV(ctx, ri, di, SpkBin2)
-	sb2 := NrnV(ctx, si, di, SpkBin2)
-	rb3 := NrnV(ctx, ri, di, SpkBin3)
-	sb3 := NrnV(ctx, si, di, SpkBin3)
-
-	b0 := 0.1 * (rb0 * sb0)
-	b1 := 0.1 * (rb1 * sb1)
-	b2 := 0.1 * (rb2 * sb2)
-	b3 := 0.1 * (rb3 * sb3)
-
-	pj.Learn.KinaseCa.FinalCa(b0, b1, b2, b3, &syCaM, &syCaP, &syCaD)
+	var syCaM, syCaP, syCaD, caUpT float32
+	switch pj.Learn.Trace.SynCa {
+	case StdSynCa:
+		caUpT = SynCaV(ctx, syni, di, CaUpT)                                 // time of last update
+		syCaM = SynCaV(ctx, syni, di, CaM)                                   // fast time scale
+		syCaP = SynCaV(ctx, syni, di, CaP)                                   // slower but still fast time scale, drives Potentiation
+		syCaD = SynCaV(ctx, syni, di, CaD)                                   // slow time scale, drives Depression (one trial = 200 cycles)
+		pj.Learn.KinaseCa.CurCa(ctx.SynCaCtr, caUpT, &syCaM, &syCaP, &syCaD) // always update, getting current Ca (just optimization)
+	case LinearSynCa:
+		rb0 := NrnV(ctx, ri, di, SpkBin0)
+		sb0 := NrnV(ctx, si, di, SpkBin0)
+		rb1 := NrnV(ctx, ri, di, SpkBin1)
+		sb1 := NrnV(ctx, si, di, SpkBin1)
+		rb2 := NrnV(ctx, ri, di, SpkBin2)
+		sb2 := NrnV(ctx, si, di, SpkBin2)
+		rb3 := NrnV(ctx, ri, di, SpkBin3)
+		sb3 := NrnV(ctx, si, di, SpkBin3)
+
+		b0 := 0.1 * (rb0 * sb0)
+		b1 := 0.1 * (rb1 * sb1)
+		b2 := 0.1 * (rb2 * sb2)
+		b3 := 0.1 * (rb3 * sb3)
+
+		pj.Learn.KinaseCa.FinalCa(b0, b1, b2, b3, &syCaM, &syCaP, &syCaD)
+	case NeurSynCa:
+		gain := float32(1.0)
+		syCaM = gain * NrnV(ctx, si, di, CaSpkM) * NrnV(ctx, ri, di, CaSpkM)
+		syCaP = gain * NrnV(ctx, si, di, CaSpkP) * NrnV(ctx, ri, di, CaSpkP)
+		syCaD = gain * NrnV(ctx, si, di, CaSpkD) * NrnV(ctx, ri, di, CaSpkD)
+	}
 
 	SetSynCaV(ctx, syni, di, CaM, syCaM)
 	SetSynCaV(ctx, syni, di, CaP, syCaP)

axon/pathparams.go

@@ -2,7 +2,6 @@ package axon
 
 import (
 	"cogentcore.org/core/vgpu/gosl/slrand"
-	"cogentcore.org/core/vgpu/gosl/sltype"
 )
 
 //gosl:hlsl axonrand
@@ -32,8 +31,7 @@ func GetRandomNumber(index uint32, counter slrand.Counter, funIndex RandFunIndex
 	var randCtr slrand.Counter
 	randCtr = counter
 	randCtr.Add(uint32(funIndex))
-	var ctr sltype.Uint2
-	ctr = randCtr.Uint2()
+	ctr := randCtr.Uint2()
 	return slrand.Float(&ctr, index)
 }
 

axon/rand.go

@@ -2,7 +2,7 @@
 # The go generate command does this automatically.
 
 all: 
-	cd ../; gosl -exclude=Update,UpdateParams,Defaults,AllParams,ShouldShow cogentcore.org/core/math32/v2/fastexp.go cogentcore.org/core/etable/v2/minmax ../chans/chans.go ../chans ../kinase ../fsfffb/inhib.go ../fsfffb github.com/emer/emergent/v2/etime github.com/emer/emergent/v2/ringidx rand.go avgmax.go neuromod.go globals.go context.go neuron.go synapse.go pool.go layervals.go act.go act_prjn.go inhib.go learn.go layertypes.go layerparams.go deep_layers.go rl_layers.go pvlv_layers.go pcore_layers.go prjntypes.go prjnparams.go deep_prjns.go rl_prjns.go pvlv_prjns.go pcore_prjns.go hip_prjns.go gpu_hlsl
+	cd ../; gosl -exclude=Update,UpdateParams,Defaults,AllParams,ShouldShow cogentcore.org/core/math32/fastexp.go cogentcore.org/core/math32/minmax ../chans/chans.go ../chans ../kinase ../fsfffb/inhib.go ../fsfffb github.com/emer/emergent/v2/etime github.com/emer/emergent/v2/ringidx rand.go avgmax.go neuromod.go globals.go context.go neuron.go synapse.go pool.go layervals.go act.go act_prjn.go inhib.go learn.go layertypes.go layerparams.go deep_layers.go rl_layers.go pvlv_layers.go pcore_layers.go prjntypes.go prjnparams.go deep_prjns.go rl_prjns.go pvlv_prjns.go pcore_prjns.go hip_prjns.go gpu_hlsl
 
 # note: gosl automatically compiles the hlsl files using this command:
 %.spv : %.hlsl