learn.go

// Code generated by "goal build"; DO NOT EDIT.
//line learn.goal:1
// Copyright (c) 2019, 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 axon

import (
	"cogentcore.org/core/math32"
	"cogentcore.org/core/math32/minmax"
	"cogentcore.org/lab/base/randx"
	"cogentcore.org/lab/gosl/slbool"
	"github.com/emer/axon/v2/chans"
	"github.com/emer/axon/v2/kinase"
)

////////  learn.go contains the learning params and functions for axon

//gosl:start
//gosl:import "github.com/emer/axon/v2/kinase"

// LearnCaParams parameterizes the neuron-level calcium signals driving learning:
// LearnCa = NMDA + VGCC Ca sources, where VGCC can be simulated from spiking or
// use the more complex and dynamaic VGCC channel directly.
// LearnCa is then integrated in a cascading manner at multiple time scales:
// CaM (as in calmodulin), CaP (ltP, CaMKII, plus phase), CaD (ltD, DAPK1, minus phase).
type LearnCaParams struct {

	// denomenator used for normalizing LearnCa, so the max is roughly 1 - 1.5 or so, which works best in terms of previous standard learning rules, and overall learning performance
	Norm float32 `default:"80"`

	// use spikes to generate VGCC instead of actual VGCC current -- see SpkVGCCa for calcium contribution from each spike
	SpkVGCC slbool.Bool `default:"true"`

	// multiplier on spike for computing Ca contribution to LearnCa in SpkVGCC mode
	SpkVgccCa float32 `default:"35"`

	// time constant of decay for VgccCa calcium -- it is highly transient around spikes, so decay and diffusion factors are more important than for long-lasting NMDA factor.  VgccCa is integrated separately int VgccCaInt prior to adding into NMDA Ca in LearnCa
	VgccTau float32 `default:"10"`

	// time constants for integrating LearnCa across M, P and D cascading levels
	Dt kinase.CaDtParams `display:"inline"`

	// Threshold on CaP CaD value for updating synapse-level Ca values (SynCa) -- this is purely a performance optimization that excludes random infrequent spikes -- 0.05 works well on larger networks but not smaller, which require the .01 default.
	UpdateThr float32 `default:"0.01,0.02,0.5"`

	// rate = 1 / tau
	VgccDt float32 `display:"-" json:"-" xml:"-" edit:"-"`

	// = 1 / Norm
	NormInv float32 `display:"-" json:"-" xml:"-" edit:"-"`

	pad int32
}

func (np *LearnCaParams) Defaults() {
	np.Norm = 80
	np.SpkVGCC.SetBool(true)
	np.SpkVgccCa = 35
	np.UpdateThr = 0.01
	np.VgccTau = 10
	np.Dt.Defaults()
	np.Dt.MTau = 2
	np.Update()
}

func (np *LearnCaParams) Update() {
	np.Dt.Update()
	np.VgccDt = 1 / np.VgccTau
	np.NormInv = 1 / np.Norm
}

// VgccCa updates the simulated VGCC calcium from spiking, if that option is selected,
// and performs time-integration of VgccCa
func (np *LearnCaParams) VgccCaFromSpike(ctx *Context, ni, di uint32) {
	if np.SpkVGCC.IsTrue() {
		Neurons.Set(np.SpkVgccCa*Neurons.Value(int(ni), int(di), int(Spike)), int(ni), int(di), int(VgccCa))
	}
	Neurons.SetAdd(Neurons.Value(int(ni), int(di), int(VgccCa))-np.VgccDt*Neurons.Value(int(ni), int(di), int(VgccCaInt)), int(ni), int(di), int(VgccCaInt))
	// Dt only affects decay, not rise time
}

// LearnCas updates the LearnCa value and its cascaded values, based on NMDA, VGCC Ca
// it first calls VgccCa to update the spike-driven version of that variable, and
// perform its time-integration.
func (np *LearnCaParams) LearnCas(ctx *Context, ni, di uint32) {
	np.VgccCaFromSpike(ctx, ni, di)
	Neurons.Set(np.NormInv*(Neurons.Value(int(ni), int(di), int(NmdaCa))+Neurons.Value(int(ni), int(di), int(VgccCaInt))), int(ni), int(di), int(LearnCa))
	Neurons.SetAdd(np.Dt.MDt*(Neurons.Value(int(ni), int(di), int(LearnCa))-Neurons.Value(int(ni), int(di), int(LearnCaM))), int(ni), int(di), int(LearnCaM))
	Neurons.SetAdd(np.Dt.PDt*(Neurons.Value(int(ni), int(di), int(LearnCaM))-Neurons.Value(int(ni), int(di), int(LearnCaP))), int(ni), int(di), int(LearnCaP))
	Neurons.SetAdd(np.Dt.DDt*(Neurons.Value(int(ni), int(di), int(LearnCaP))-Neurons.Value(int(ni), int(di), int(LearnCaD))), int(ni), int(di), int(LearnCaD))
	Neurons.Set(Neurons.Value(int(ni), int(di), int(LearnCaP))-Neurons.Value(int(ni), int(di), int(LearnCaD)), int(ni), int(di), int(CaDiff))
}

////////  GateSyncParams

// GateSyncParams enable gating timing in another layer to synchronize the synaptic
// calcium learning signals in this layer.
type GateSyncParams struct {
	// On activates this mechanism.
	On slbool.Bool

	// Offset is time offset from gating signal to effects on spike accumulation.
	Offset int32

	// GateLayIndex is the index of the Layer to get gating timing from;
	// set during Build from BuildConfig GateLayName if present: -1 if not used.
	GateLayIndex int32 `edit:"-"`

	pad int32
}

func (gs *GateSyncParams) Defaults() {
	gs.Offset = 50
}

func (gs *GateSyncParams) Update() {
}

func (gs *GateSyncParams) ShouldDisplay(field string) bool {
	switch field {
	case "On":
		return true
	default:
		return gs.On.IsTrue()
	}
}

//
// func (gs *GateSyncParams) ShiftBins(delta, minusBins, plusBins int32, ni, di uint32) {
// 	if delta > 0 { // later than typical: shift left
// 		gs.ShiftToEarlier(delta, minusBins, plusBins, ni, di)
// 	} else {
// 		gs.ShiftToLater(delta, minusBins, plusBins, ni, di)
// 	}
// 	for i := range plusBins {
// 		Neurons[ni, di, SpikeBin0 + NeuronVars(minusBins+i)] = 0.0
// 	}
// }
//
// // todo: linear weighting of neighboring 2 values
//
// func (gs *GateSyncParams) ShiftToEarlier(delta, minusBins, plusBins int32, ni, di uint32) {
// 	del := min(delta, plusBins)
// 	for i := range minusBins {
// 		Neurons[ni, di, SpikeBin0 + NeuronVars(i)] = Neurons[ni, di, SpikeBin0 + NeuronVars(i+del)]
// 	}
// }
//
// func (gs *GateSyncParams) ShiftToLater(delta, minusBins, plusBins int32, ni, di uint32) {
// 	del := min(-delta, plusBins)
// 	for i := minusBins-1; i>=del; i-- {
// 		Neurons[ni, di, SpikeBin0 + NeuronVars(i)] = Neurons[ni, di, SpikeBin0 + NeuronVars(i-del)]
// 	}
// 	fill := Neurons[ni, di, SpikeBin0 + NeuronVars(del)]
// 	for i := range del {
// 		Neurons[ni, di, SpikeBin0 + NeuronVars(i)] = fill
// 	}
// }

////////  TrgAvgActParams

// TrgAvgActParams govern the target and actual long-term average activity in neurons.
// Target value is adapted by neuron-wise error and difference in actual vs. target.
// drives synaptic scaling at a slow timescale (Network.SlowInterval).
type TrgAvgActParams struct {

	// if this is > 0, then each neuron's GiBase is initialized as this proportion of TrgRange.Max - TrgAvg -- gives neurons differences in intrinsic inhibition / leak as a starting bias.  This is independent of using the target values to scale synaptic weights.
	GiBaseInit float32

	// whether to use target average activity mechanism to rescale synaptic weights, so that activity tracks the target values
	RescaleOn slbool.Bool

	// learning rate for adjustments to Trg value based on unit-level error signal.  Population TrgAvg values are renormalized to fixed overall average in TrgRange. Generally, deviating from the default doesn't make much difference.
	ErrLRate float32 `default:"0.02"`

	// rate parameter for how much to scale synaptic weights in proportion to the AvgDif between target and actual proportion activity -- this determines the effective strength of the constraint, and larger models may need more than the weaker default value.
	SynScaleRate float32 `default:"0.005,0.0002"`

	// amount of mean trg change to subtract -- 1 = full zero sum.  1 works best in general -- but in some cases it may be better to start with 0 and then increase using network SetSubMean method at a later point.
	SubMean float32 `default:"0,1"`

	// permute the order of TrgAvg values within layer -- otherwise they are just assigned in order from highest to lowest for easy visualization -- generally must be true if any topographic weights are being used
	Permute slbool.Bool `default:"true"`

	// use pool-level target values if pool-level inhibition and 4D pooled layers are present -- if pool sizes are relatively small, then may not be useful to distribute targets just within pool
	Pool slbool.Bool

	pad int32

	// range of target normalized average activations -- individual neurons are assigned values within this range to TrgAvg, and clamped within this range.
	TrgRange minmax.F32 `default:"{'Min':0.5,'Max':2}"`
}

func (ta *TrgAvgActParams) Update() {
}

func (ta *TrgAvgActParams) Defaults() {
	ta.RescaleOn.SetBool(true)
	ta.ErrLRate = 0.02
	ta.SynScaleRate = 0.005
	ta.SubMean = 1 // 1 in general beneficial
	ta.TrgRange.Set(0.5, 2)
	ta.Permute.SetBool(true)
	ta.Pool.SetBool(true)
	ta.Update()
}

func (ta *TrgAvgActParams) ShouldDisplay(field string) bool {
	switch field {
	case "RescaleOn", "GiBaseInit":
		return true
	case "TrgRange":
		return ta.RescaleOn.IsTrue() || ta.GiBaseInit > 0
	default:
		return ta.RescaleOn.IsTrue()
	}
}

////////  RLRateParams

// RLRateParams are recv neuron learning rate modulation parameters.
// Has two factors: the derivative of the sigmoid based on CaD
// activity levels, and based on the phase-wise differences in activity (Diff).
type RLRateParams struct {

	// use learning rate modulation
	On slbool.Bool `default:"true"`

	// use a linear sigmoid function: if act > .5: 1-act; else act
	// otherwise use the actual sigmoid derivative which is squared: a(1-a)
	SigmoidLinear slbool.Bool `default:"true"`

	// minimum learning rate multiplier for sigmoidal act (1-act) factor,
	// which prevents lrate from going too low for extreme values.
	// Set to 1 to disable Sigmoid derivative factor, which is default for Target layers.
	SigmoidMin float32 `default:"0.05,1"`

	// modulate learning rate as a function of plus - minus differences
	Diff slbool.Bool

	// threshold on Max(CaP, CaD) below which Min lrate applies.
	// must be > 0 to prevent div by zero.
	SpkThr float32 `default:"0.1"`

	// threshold on recv neuron error delta, i.e., |CaP - CaD| below which lrate is at Min value
	DiffThr float32 `default:"0.02"`

	// for Diff component, minimum learning rate value when below ActDiffThr
	Min float32 `default:"0.001"`

	pad int32
}

func (rl *RLRateParams) Update() {
}

func (rl *RLRateParams) Defaults() {
	rl.On.SetBool(true)
	rl.SigmoidLinear.SetBool(true)
	rl.SigmoidMin = 0.05
	rl.Diff.SetBool(true)
	rl.SpkThr = 0.1
	rl.DiffThr = 0.02
	rl.Min = 0.001
	rl.Update()
}

func (rl *RLRateParams) ShouldDisplay(field string) bool {
	switch field {
	case "On":
		return true
	case "Diff", "SigmoidMin", "SigmoidLinear":
		return rl.On.IsTrue()
	default:
		return rl.On.IsTrue() && rl.Diff.IsTrue()
	}
}

// RLRateSigDeriv returns the sigmoid derivative learning rate
// factor as a function of spiking activity, with mid-range values having
// full learning and extreme values a reduced learning rate:
// deriv = 4*act*(1-act) or linear: if act > .5: 2*(1-act); else 2*act
// The activity should be CaP and the layer maximum is used
// to normalize that to a 0-1 range.
func (rl *RLRateParams) RLRateSigDeriv(act float32, laymax float32) float32 {
	if rl.On.IsFalse() || laymax == 0 {
		return 1.0
	}
	ca := min(act/laymax, 1.0)
	var lr float32
	if rl.SigmoidLinear.IsTrue() {
		if ca < 0.5 {
			lr = 2 * ca
		} else {
			lr = 2 * (1 - ca)
		}
	} else {
		lr = 4.0 * ca * (1 - ca) // .5 * .5 = .25 = peak
	}
	if lr < rl.SigmoidMin {
		lr = rl.SigmoidMin
	}
	return lr
}

// RLRateDiff returns the learning rate as a function of difference between
// CaP and CaD values
func (rl *RLRateParams) RLRateDiff(scap, scad float32) float32 {
	if rl.On.IsFalse() || rl.Diff.IsFalse() {
		return 1.0
	}
	smax := math32.Max(scap, scad)
	if smax > rl.SpkThr { // avoid div by 0
		dif := math32.Abs(scap - scad)
		if dif < rl.DiffThr {
			return rl.Min
		}
		return (dif / smax)
	}
	return rl.Min
}

// LearnNeuronParams manages learning-related parameters at the neuron-level.
// This is mainly the running average activations that drive learning
type LearnNeuronParams struct {

	// CaLearn parameterizes the neuron-level calcium signals driving learning:
	// LearnCa = NMDA + VGCC Ca sources, where VGCC can be simulated from spiking
	// or use the more complex and dynamic VGCC channel directly.  LearnCa is then
	// integrated in a cascading manner at multiple time scales:
	// LearnCaM (as in calmodulin), LearnCaP (ltP, CaMKII, plus phase),
	// LearnCaD (ltD, DAPK1, minus phase).
	CaLearn LearnCaParams `display:"inline"`

	// CaSpike parameterizes the neuron-level spike-driven calcium signals:
	// CaM (calmodulin), CaP (ltP, CaMKII, plus phase), CaD (ltD, DAPK1, minus phase).
	// These values are used in various cases as a proxy for the activation (spiking)
	// based learning signal.
	CaSpike kinase.CaSpikeParams `display:"inline"`

	// NMDA channel parameters used for learning, vs. the ones driving activation.
	// This allows exploration of learning parameters independent of their effects
	// on active maintenance contributions of NMDA, and may be supported by different
	// receptor subtypes.
	LearnNMDA chans.NMDAParams `display:"inline"`

	// GateSync enables gating timing in another layer to synchronize the synaptic
	// calcium learning signals in this layer.
	GateSync GateSyncParams `display:"inline"`

	// TrgAvgAct has the synaptic scaling parameters for regulating overall average
	// activity compared to neuron's own target level.
	TrgAvgAct TrgAvgActParams `display:"inline"`

	// RLRate has the recv neuron learning rate modulation params: an additional
	// error-based modulation of learning for receiver side:
	// RLRate = |CaP - CaD| / Max(CaP, CaD)
	RLRate RLRateParams `display:"inline"`

	// NeuroMod parameterizes neuromodulation effects on learning rate and activity,
	// as a function of layer-level DA and ACh values, which are updated from global
	// Context values, and computed from reinforcement learning algorithms.
	NeuroMod NeuroModParams `display:"inline"`
}

func (ln *LearnNeuronParams) Update() {
	ln.CaLearn.Update()
	ln.CaSpike.Update()
	ln.LearnNMDA.Update()
	ln.GateSync.Update()
	ln.TrgAvgAct.Update()
	ln.RLRate.Update()
	ln.NeuroMod.Update()
}

func (ln *LearnNeuronParams) Defaults() {
	ln.CaLearn.Defaults()
	ln.CaSpike.Defaults()
	ln.LearnNMDA.Defaults()
	ln.LearnNMDA.ITau = 1
	ln.LearnNMDA.Update()
	ln.GateSync.Defaults()
	ln.TrgAvgAct.Defaults()
	ln.RLRate.Defaults()
	ln.NeuroMod.Defaults()
}

// InitLearnNeurCa initializes the neuron-level calcium learning and spking variables.
// Called by InitWeights (at start of learning).
func (ln *LearnNeuronParams) InitNeurCa(ctx *Context, ni, di uint32) {
	Neurons.Set(0, int(ni), int(di), int(GnmdaLrn))
	Neurons.Set(0, int(ni), int(di), int(NmdaCa))

	Neurons.Set(0, int(ni), int(di), int(VgccCa))
	Neurons.Set(0, int(ni), int(di), int(VgccCaInt))

	Neurons.Set(0, int(ni), int(di), int(LearnCa))

	Neurons.Set(0, int(ni), int(di), int(CaM))
	Neurons.Set(0, int(ni), int(di), int(CaP))
	Neurons.Set(0, int(ni), int(di), int(CaD))

	Neurons.Set(0, int(ni), int(di), int(LearnCaM))
	Neurons.Set(0, int(ni), int(di), int(LearnCaP))
	Neurons.Set(0, int(ni), int(di), int(LearnCaD))
	Neurons.Set(0, int(ni), int(di), int(CaDiff))
}

// LearnNMDAFromRaw updates the separate NMDA conductance and calcium values
// based on GeTot = GeRaw + external ge conductance.  These are the variables
// that drive learning -- can be the same as activation but also can be different
// for testing learning Ca effects independent of activation effects.
func (ln *LearnNeuronParams) LearnNMDAFromRaw(ctx *Context, ni, di uint32, geTot float32) {
	geEff := max(geTot, 0.0)
	vmd := Neurons.Value(int(ni), int(di), int(VmDend))
	Neurons.Set(ln.LearnNMDA.NMDASyn(Neurons.Value(int(ni), int(di), int(GnmdaLrn)), geEff), int(ni), int(di), int(GnmdaLrn))
	gnmda := ln.LearnNMDA.Gnmda(Neurons.Value(int(ni), int(di), int(GnmdaLrn)), vmd)
	Neurons.Set(float32(gnmda*ln.LearnNMDA.CaFromV(vmd)), int(ni), int(di), int(NmdaCa))
}

// CaFromSpike updates all spike-driven calcium variables, including LearnCa and CaSpike.
// Computed after new activation for current cycle is updated.
func (ln *LearnNeuronParams) CaFromSpike(ctx *Context, ni, di uint32) {
	caSpkM := Neurons.Value(int(ni), int(di), int(CaM))
	caSpkP := Neurons.Value(int(ni), int(di), int(CaP))
	caSpkD := Neurons.Value(int(ni), int(di), int(CaD))
	ln.CaSpike.CaFromSpike(Neurons.Value(int(ni), int(di), int(Spike)), &caSpkM, &caSpkP, &caSpkD)
	Neurons.Set(caSpkM, int(ni), int(di), int(CaM))
	Neurons.Set(caSpkP, int(ni), int(di), int(CaP))
	Neurons.Set(caSpkD, int(ni), int(di), int(CaD))

	ln.CaLearn.LearnCas(ctx, ni, di)
}

////////  SWtParams

// SigFun is the sigmoid function for value w in 0-1 range, with gain and offset params
func SigFun(w, gain, off float32) float32 {
	if w <= 0 {
		return 0
	}
	if w >= 1 {
		return 1
	}
	return (1 / (1 + math32.Pow((off*(1-w))/w, gain)))
}

// SigFun61 is the sigmoid function for value w in 0-1 range, with default gain = 6, offset = 1 params
func SigFun61(w float32) float32 {
	if w <= 0 {
		return 0
	}
	if w >= 1 {
		return 1
	}
	pw := (1 - w) / w
	return (1 / (1 + pw*pw*pw*pw*pw*pw))
}

// SigInvFun is the inverse of the sigmoid function
func SigInvFun(w, gain, off float32) float32 {
	if w <= 0 {
		return 0
	}
	if w >= 1 {
		return 1
	}
	return 1.0 / (1.0 + math32.Pow((1.0-w)/w, 1/gain)/off)
}

// SigInvFun61 is the inverse of the sigmoid function, with default gain = 6, offset = 1 params
func SigInvFun61(w float32) float32 {
	if w <= 0 {
		return 0
	}
	if w >= 1 {
		return 1
	}
	rval := 1.0 / (1.0 + math32.Pow((1.0-w)/w, 1.0/6.0))
	return rval
}

// SWtInitParams for initial SWt values
type SWtInitParams struct {

	// how much of the initial random weights are captured in the SWt values -- rest goes into the LWt values.  1 gives the strongest initial biasing effect, for larger models that need more structural support. 0.5 should work for most models where stronger constraints are not needed.
	SPct float32 `min:"0" max:"1" default:"0,1,0.5"`

	// target mean weight values across receiving neuron's pathway -- the mean SWt values are constrained to remain at this value.  some pathways may benefit from lower mean of .4
	Mean float32 `default:"0.5,0.4"`

	// initial variance in weight values, prior to constraints.
	Var float32 `default:"0.25"`

	// symmetrize the initial weight values with those in reciprocal pathway -- typically true for bidirectional excitatory connections
	Sym slbool.Bool `default:"true"`
}

func (sp *SWtInitParams) Defaults() {
	sp.SPct = 0.5
	sp.Mean = 0.5
	sp.Var = 0.25
	sp.Sym.SetBool(true)
}

func (sp *SWtInitParams) Update() {
}

// SWtAdaptParams manages adaptation of SWt values
type SWtAdaptParams struct {

	// if true, adaptation is active -- if false, SWt values are not updated, in which case it is generally good to have Init.SPct=0 too.
	On slbool.Bool

	// learning rate multiplier on the accumulated DWt values (which already have fast LRate applied) to incorporate into SWt during slow outer loop updating -- lower values impose stronger constraints, for larger networks that need more structural support, e.g., 0.001 is better after 1,000 epochs in large models.  0.1 is fine for smaller models.
	LRate float32 `default:"0.1,0.01,0.001,0.0002"`

	// amount of mean to subtract from SWt delta when updating -- generally best to set to 1
	SubMean float32 `default:"1"`

	// gain of sigmoidal constrast enhancement function used to transform learned, linear LWt values into Wt values
	SigGain float32 `default:"6"`
}

func (sp *SWtAdaptParams) Defaults() {
	sp.On.SetBool(true)
	sp.LRate = 0.1
	sp.SubMean = 1
	sp.SigGain = 6
	sp.Update()
}

func (sp *SWtAdaptParams) Update() {
}

func (sp *SWtAdaptParams) ShouldDisplay(field string) bool {
	switch field {
	case "On":
		return true
	default:
		return sp.On.IsTrue()
	}
}

//gosl:end

// RandVar returns the random variance in weight value (zero mean) based on Var param
func (sp *SWtInitParams) RandVar(rnd randx.Rand) float32 {
	return sp.Var * 2.0 * (rnd.Float32() - 0.5)
}

// // RandVar returns the random variance (zero mean) based on DreamVar param
// func (sp *SWtAdaptParams) RandVar(rnd randx.Rand) float32 {
// 	return sp.DreamVar * 2.0 * (rnd.Float32(-1) - 0.5)
// }

//gosl:start

// SWtParams manages structural, slowly adapting weight values (SWt),
// in terms of initialization and updating over course of learning.
// SWts impose initial and slowly adapting constraints on neuron connectivity
// to encourage differentiation of neuron representations and overall good behavior
// in terms of not hogging the representational space.
// The TrgAvg activity constraint is not enforced through SWt -- it needs to be
// more dynamic and supported by the regular learned weights.
type SWtParams struct {

	// initialization of SWt values
	Init SWtInitParams `display:"inline"`

	// adaptation of SWt values in response to LWt learning
	Adapt SWtAdaptParams `display:"inline"`

	// range limits for SWt values
	Limit minmax.F32 `default:"{'Min':0.2,'Max':0.8}" display:"inline"`
}

func (sp *SWtParams) Defaults() {
	sp.Init.Defaults()
	sp.Adapt.Defaults()
	sp.Limit.Set(0.2, 0.8)
}

func (sp *SWtParams) Update() {
	sp.Init.Update()
	sp.Adapt.Update()
}

// WtVal returns the effective Wt value given the SWt and LWt values
func (sp *SWtParams) WtValue(swt, lwt float32) float32 {
	return swt * sp.SigFromLinWt(lwt)
}

// ClipSWt returns SWt value clipped to valid range
func (sp *SWtParams) ClipSWt(swt float32) float32 {
	return sp.Limit.ClampValue(swt)
}

// ClipWt returns Wt value clipped to 0-1 range
func (sp *SWtParams) ClipWt(wt float32) float32 {
	if wt > 1 {
		return 1
	}
	if wt < 0 {
		return 0
	}
	return wt
}

// SigFromLinWt returns sigmoidal contrast-enhanced weight from linear weight,
// centered at 1 and normed in range +/- 1 around that
// in preparation for multiplying times SWt
func (sp *SWtParams) SigFromLinWt(lw float32) float32 {
	var wt float32
	if sp.Adapt.SigGain == 1 {
		wt = lw
	} else if sp.Adapt.SigGain == 6 {
		wt = SigFun61(lw)
	} else {
		wt = SigFun(lw, sp.Adapt.SigGain, 1)
	}
	return 2.0 * wt // center at 1 instead of .5
}

// LinFromSigWt returns linear weight from sigmoidal contrast-enhanced weight.
// wt is centered at 1, and normed in range +/- 1 around that,
// return value is in 0-1 range, centered at .5
func (sp *SWtParams) LinFromSigWt(wt float32) float32 {
	wte := wt * 0.5
	if wte < 0 {
		wte = 0
	} else if wte > 1 {
		wte = 1
	}
	if sp.Adapt.SigGain == 1 {
		return wte
	}
	if sp.Adapt.SigGain == 6 {
		return SigInvFun61(wte)
	}
	return SigInvFun(wte, sp.Adapt.SigGain, 1)
}

// LWtFromWts returns linear, learning LWt from wt and swt.
// LWt is set to reproduce given Wt relative to given SWt base value.
func (sp *SWtParams) LWtFromWts(wt, swt float32) float32 {
	rwt := wt / swt
	return sp.LinFromSigWt(rwt)
}

// WtFromDWt updates the synaptic weights from accumulated weight changes.
// wt is the sigmoidal contrast-enhanced weight and lwt is the linear weight value.
func (sp *SWtParams) WtFromDWt(wt, lwt *float32, dwt, swt float32) {
	if dwt == 0 {
		if *wt == 0 { // restore failed wts
			*wt = sp.WtValue(swt, *lwt)
		}
		return
	}
	// note: softbound happened at dwt stage
	*lwt += dwt
	if *lwt < 0 {
		*lwt = 0
	} else if *lwt > 1 {
		*lwt = 1
	}
	*wt = sp.WtValue(swt, *lwt)
}

//gosl:end

// InitWeightsSyn initializes weight values based on WtInit randomness parameters
// for an individual synapse.
// It also updates the linear weight value based on the sigmoidal weight value.
func (sp *SWtParams) InitWeightsSyn(ctx *Context, syni uint32, rnd randx.Rand, mean, spct float32) {
	wtv := sp.Init.RandVar(rnd)
	wt := mean + wtv
	Synapses.Set(wt, int(syni), int(Wt))
	Synapses.Set(sp.ClipSWt(mean+spct*wtv), int(syni), int(SWt))
	if spct == 0 { // this is critical for weak init wt, SPCt = 0 paths
		Synapses.Set(0.5, int(syni), int(SWt))
	}
	Synapses.Set(sp.LWtFromWts(wt, Synapses.Value(int(syni), int(SWt))), int(syni), int(LWt))
	Synapses.Set(0, int(syni), int(DWt))
	Synapses.Set(0, int(syni), int(DSWt))
}

// InitWeightsSynTrace initializes SynapseTrace values
// for an individual synapse.
func (sp *SWtParams) InitWeightsSynTrace(ctx *Context, syni, di uint32) {
	SynapseTraces.Set(0, int(syni), int(di), int(Tr))
	SynapseTraces.Set(0, int(syni), int(di), int(DTr))
	SynapseTraces.Set(0, int(syni), int(di), int(DiDWt))
}

//gosl:start

// LRateParams manages learning rate parameters
type LRateParams struct {

	// base learning rate for this pathway -- can be modulated by other factors below -- for larger networks, use slower rates such as 0.04, smaller networks can use faster 0.2.
	Base float32 `default:"0.04,0.1,0.2"`

	// scheduled learning rate multiplier, simulating reduction in plasticity over aging
	Sched float32

	// dynamic learning rate modulation due to neuromodulatory or other such factors
	Mod float32

	// effective actual learning rate multiplier used in computing DWt: Eff = eMod * Sched * Base
	Eff float32 `edit:"-"`
}

func (ls *LRateParams) Defaults() {
	ls.Base = 0.04
	ls.Sched = 1
	ls.Mod = 1
	ls.Update()
}

func (ls *LRateParams) Update() {
	ls.UpdateEff()
}

func (ls *LRateParams) UpdateEff() {
	ls.Eff = ls.Mod * ls.Sched * ls.Base
}

// Init initializes modulation values back to 1 and updates Eff
func (ls *LRateParams) Init() {
	ls.Sched = 1
	ls.Mod = 1
	ls.UpdateEff()
}

// TraceParams manages parameters associated with temporal trace learning
type TraceParams struct {

	// Tau is the time constant for integrating the synaptic trace over theta cycle timescales.
	// This governs the decay rate of syanptic trace.
	Tau float32 `default:"1,2,4"`

	// CaGain is a multiplier on the total synaptic calcium values, computed from products
	// of the neuron-level [SpikeBins] values. If [Context.SpikeBinCycles] is lower than
	// default of 25, then this value may need to be set higher.
	CaGain float32 `default:"1"`

	// SubMean is the amount of the mean dWt to subtract, producing a zero-sum effect.
	// 1.0 = full zero-sum dWt. Only applies to non-zero DWts.
	// Typically set to 0 for standard trace learning pathways, although some require it
	// for stability over the long haul. Can use SetSubMean to set to 1 after significant
	// early learning has occurred with 0. Some special path types (e.g., Hebb) benefit
	// from SubMean = 1 always.
	SubMean float32 `default:"0,1"`

	// LearnThr is the threshold for learning, for specialized learning algorithms.
	// This is not relevant for the standard kinase error-driven cortical learning algorithm.
	// In Matrix and VSPatch it applies to normalized GeIntNorm value: setting this relatively
	// high encourages sparser representations.
	LearnThr float32

	// rate = 1 / tau
	Dt float32 `display:"-" json:"-" xml:"-" edit:"-"`

	pad, pad1, pad2 float32
}

func (tp *TraceParams) Defaults() {
	tp.Tau = 1
	tp.CaGain = 1
	tp.SubMean = 0
	tp.LearnThr = 0
	tp.Update()
}

func (tp *TraceParams) Update() {
	tp.Dt = 1.0 / tp.Tau
}

// TrFromCa returns updated trace factor as function of a
// synaptic calcium update factor and current trace
func (tp *TraceParams) TrFromCa(tr float32, ca float32) float32 {
	return tr + tp.Dt*(ca-tr)
}

////////  LRateMod

// LRateMod implements global learning rate modulation, based on a performance-based
// factor, for example error.  Increasing levels of the factor = higher learning rate.
// This can be added to a Sim and called prior to DWt() to dynamically change lrate
// based on overall network performance.
type LRateMod struct {

	// toggle use of this modulation factor
	On slbool.Bool

	// baseline learning rate -- what you get for correct cases
	Base float32 `min:"0" max:"1"`

	pad, pad1 int32

	// defines the range over which modulation occurs for the modulator factor -- Min and below get the Base level of learning rate modulation, Max and above get a modulation of 1
	Range minmax.F32
}

func (lr *LRateMod) Defaults() {
	lr.On.SetBool(true)
	lr.Base = 0.2
	lr.Range.Set(0.2, 0.8)
}

func (lr *LRateMod) Update() {
}

func (lr *LRateMod) ShouldDisplay(field string) bool {
	switch field {
	case "On":
		return true
	default:
		return lr.On.IsTrue()
	}
}

// Mod returns the learning rate modulation factor as a function
// of any kind of normalized modulation factor, e.g., an error measure.
// If fact <= Range.Min, returns Base
// If fact >= Range.Max, returns 1
// otherwise, returns proportional value between Base..1
func (lr *LRateMod) Mod(fact float32) float32 {
	lrm := lr.Range.NormValue(fact) // clips to 0-1 range
	md := lr.Base + lrm*(1-lr.Base) // resulting mod is in Base-1 range
	return md
}

//gosl:end

// LRateMod calls LRateMod on given network, using computed Mod factor
// based on given normalized modulation factor
// (0 = no error = Base learning rate, 1 = maximum error).
// returns modulation factor applied.
func (lr *LRateMod) LRateMod(net *Network, fact float32) float32 {
	if lr.Range.Max == 0 {
		lr.Defaults()
	}
	if lr.On.IsFalse() {
		return 1
	}
	md := lr.Mod(fact)
	net.LRateMod(md)
	return md
}

//gosl:start

////////  HebbParams

// HebbParams for optional hebbian learning that replaces the
// default learning rule, based on S = sending activity,
// R = receiving activity
type HebbParams struct {

	// if On, then the standard learning rule is replaced with a Hebbian learning rule
	On slbool.Bool

	// strength multiplier for hebbian increases, based on R * S * (1-LWt)
	Up float32 `default:"0.5"`

	// strength multiplier for hebbian decreases, based on R * (1 - S) * LWt
	Down float32 `default:"1"`

	pad float32
}

func (hp *HebbParams) Defaults() {
	hp.Up = 0.5
	hp.Down = 1
}

func (hp *HebbParams) Update() {
}

func (hp *HebbParams) ShouldDisplay(field string) bool {
	switch field {
	case "On":
		return true
	default:
		return hp.On.IsTrue()
	}
}

////////  LearnSynParams

// LearnSynParams manages learning-related parameters at the synapse-level.
type LearnSynParams struct {

	// enable learning for this pathway
	Learn slbool.Bool

	pad, pad1, pad2 int32

	// learning rate parameters, supporting two levels of modulation on top of base learning rate.
	LRate LRateParams `display:"inline"`

	// trace-based learning parameters
	Trace TraceParams `display:"inline"`

	// hebbian learning option, which overrides the default learning rules
	Hebb HebbParams `display:"inline"`
}

func (ls *LearnSynParams) Update() {
	ls.LRate.Update()
	ls.Trace.Update()
	ls.Hebb.Update()
}

func (ls *LearnSynParams) Defaults() {
	ls.Learn.SetBool(true)
	ls.LRate.Defaults()
	ls.Trace.Defaults()
	ls.Hebb.Defaults()
}

func (ls *LearnSynParams) ShouldDisplay(field string) bool {
	switch field {
	case "Learn":
		return true
	default:
		return ls.Learn.IsTrue()
	}
}

// CHLdWt returns the error-driven weight change component for a
// CHL contrastive hebbian learning rule, optionally using the checkmark
// temporally eXtended Contrastive Attractor Learning (XCAL) function
func (ls *LearnSynParams) CHLdWt(suCaP, suCaD, ruCaP, ruCaD float32) float32 {
	srp := suCaP * ruCaP
	srd := suCaD * ruCaD
	return srp - srd
}

// DeltaDWt returns the error-driven weight change component for a
// simple delta between a minus and plus phase factor, optionally using the checkmark
// temporally eXtended Contrastive Attractor Learning (XCAL) function
func (ls *LearnSynParams) DeltaDWt(plus, minus float32) float32 {
	return plus - minus
}

//gosl:end