const double Neuron::get_delta(double avg_activation)
{
static double sparsity_penalty_weight = 3;
static double sparsity = .05;
double delta = 0;
for(Synapse *out_synapse : outgoing_synapses) {
delta += out_synapse->weight * out_synapse->to->delta;
}
//printf("%f\n", avg_activation);
/*
delta += sparsity_penalty_weight *
(-(sparsity / avg_activation) + ((1 - sparsity) / (1 - avg_activation)));
*/
delta *= activation_derivative();
return delta;
}
void MultilayerPerceptron::backpropagation(cv::Mat expected,
cv::Mat actual)
{
cv::Mat error = actual - expected;
float* error_input_ptr = error.ptr< float >(0);
float* error_output_ptr = _error_signals.ptr< float >(
_error_signals.rows - 1);
for(int i = 0; i < error.cols; ++i)
{
error_output_ptr[i] = error_input_ptr[i];
}
int* layer_ptr = _layers.ptr< int >(0);
layer_ptr++; //Skip input layer;
for(int l = _weights.size() - 1; l >= 0; --l)
{
float* sum_ptr = _last_sums.ptr< float >(l);
float* gradient_ptr = _last_gradients.ptr< float >(l);
error_input_ptr = _error_signals.ptr< float >(l + 1);
error_output_ptr = _error_signals.ptr< float >(l);
//Clear the error signal from the previous layer
for(int n = 0; n < layer_ptr[l - 1]; ++n)
{
error_output_ptr[n] = 0.0f;
}
for(int n = 0; n < layer_ptr[l]; ++n)
{
float* weight_ptr = _weights[l].ptr< float >(n);
gradient_ptr[n] = error_input_ptr[n]
* activation_derivative(sum_ptr[n]);
//Use number of nodes of previous layer as number of inputs.
for(int w = 0; w < layer_ptr[l - 1]; ++w)
{
error_output_ptr[w] += gradient_ptr[n] * weight_ptr[w];
}
}
}
}
const double Neuron::get_delta_for_label(const double label)
{
return (label - activation) * activation_derivative();
}