org.kramerlab.autoencoder.neuralnet
Returns the gradient (Layer-valued) and the backpropagated error, which is passed to the layer below.
Returns the gradient (Layer-valued) and the backpropagated error, which is passed to the layer below.
This method can rely on the fact that the propagate method
already has been called in the first pass.
error propagated from above, formatted the same way (one row for each example) as input and output
gradient (Layer-valued) and the next backpropagated error
Returns the output given the input.
Returns the output given the input. This method can cache data that could be useful on the second pass of the backpropagation.
The input contains one example in each row, the output shall have the same layout.
Creates a new independent layer that has the same type as this one, but propagates the information in reverse direction
Returns the result of signal propagation in reverse direction
Color map for the activities
Optionally, one can specify how to reshape the neuron activities for visualization (height, width).
Represents a single layer of a neural network. The sort of layer meant here does not necessary contain any neurons, instead we think of a layer as of a filter in a long pipe: it can contain neurons that are activated with some activation function, but it can also contain only connections between the layer below and the layer above.
Common to all layers is that they are parameterized by something that is isomorphic to Real^n, that they know how to transform their input into output (or transmit the signal in opposite direction), and to calculate entries of the gradient of the error function that correspond to their parameters, given partial derivatives wrt. their output passed from above. With other words: each layer knows how to propagate all signals in feed-forward manner, and how to propagate errors backwards.