Cache return values for each layer at fit mode.

[I159]

Motivation

Back propagation requires return values of each layer due to a common weights correction algorithm:

w_ij -= C(A(x_ij), d) * x_ij (For each neuron)

Where:

w_ij is synapse (weight) between layer_i and layer_i-1

C() is a cost function

A() is an activation function

d is a desired output value (fit value)

x_ij is a current input value for a node

Task

Implement a "fit mode" for prediction algorithm to keep x₀₀ ... x_nn all return value vectors till back propagation be finished.

[I159]

Inputs must be cached as 2D matrix immediately after multiplication of previous layer output by weights to keep each input of each neuron for backpropagation.