forward forward

Instead of storing gradients at each layer, perform the following at each layer:

• maximize the activations for the positive examples
• minimize the activations for the negative examples

Note that gradient descent is used per-layer. But backprop is not used on the whole network