ParameterProvider has the push method that provides activations and error gradient vectors, and parameter updates for each layer are generated in LocalNeuralNetParameterProvider. Even though the methods to generate parameter updates are different by a type of layer, LocalNeuralNetParameterProvider uses one method for fully connected layer.
Thus, in order to support various layer easily, we can generate parameter updates at each layer and push them to parameter provider.
Admittedly, sending activations and error gradients can reduce network costs. But, we can do this later.
ParameterProvider
has thepush
method that provides activations and error gradient vectors, and parameter updates for each layer are generated inLocalNeuralNetParameterProvider
. Even though the methods to generate parameter updates are different by a type of layer,LocalNeuralNetParameterProvider
uses one method for fully connected layer. Thus, in order to support various layer easily, we can generate parameter updates at each layer and push them to parameter provider. Admittedly, sending activations and error gradients can reduce network costs. But, we can do this later.