public SlimRnnNeuronWithWeight(SlimRnnNeuron source, SlimRnnNeuron target, float weight, uint weightIndex, bool isEligable = false)
{
this.source = source;
this.target = target;
this.weight = weight;
this.weightIndex = weightIndex;
this.isEligable = isEligable;
}
// propagate the signal for the "next" activation
void propagate(
IList <SlimRnnNeuron> old,
IList <SlimRnnNeuron> @new,
IList <SlimRnnNeuronWithWeight> trace,
ref double time,
out bool gotoHaltLoop
)
{
gotoHaltLoop = false;
// TODO PARALLELIZE
foreach (SlimRnnNeuron u_l in old)
{
int outNeuronWithWeightIndex = -1; // set to -1 because the incrementing is done at the head of the loop
foreach (SlimRnnNeuronWithWeight w_lk in u_l.outNeuronsWithWeights)
{
outNeuronWithWeightIndex++;
// see Procedure 2.1: Spread in Schmidhuber paper
// here we invoke the learning algorithm if the connection was not jet used
if (
learningAlgorithm != null &&
!w_lk.wasUsed &&
w_lk.isEligable // call the learning algorithm only if the weight has been marked to be changable by the learning algorithm
)
{
learningAlgorithm.opportunityToAdjustWeight(w_lk);
}
w_lk.wasUsed = true;
if (w_lk.weight == 0.0f)
{
continue;
}
SlimRnnNeuron k = w_lk.target;
// add connection to trace
if (!w_lk.isMark)
{
w_lk.isMark = true;
trace.Add(w_lk);
}
if (k.type == SlimRnnNeuron.EnumType.ADDITIVE)
{
k.next += (w_lk.source.activation * w_lk.weight);
}
else
{
Debug.Assert(k.type == SlimRnnNeuron.EnumType.MULTIPLICATIVE);
k.next *= (w_lk.source.activation * w_lk.weight);
}
if (!k.used)
{
k.used = true;
@new.Add(k);
}
time += w_lk.cost; // after schmidhuber: long wires may cost more
if (time > t_lim)
{
gotoHaltLoop = true;
return;
}
}
}
if (learningAlgorithm != null && learningAlgorithm.checkForceTermination())
{
gotoHaltLoop = true;
return;
}
}
