Waffles 1.1 will include new systems for creating neural networks, which will dramatically break old versions. This will not be backwards compatible.

Features:
Removed optimizer logic from the neural network.
Combined the idea of layers and activation functions.

This will enable a user to change the optimizer of a neural network without having to worry about adapting neural network code. Changing from Stochastic Gradient Descent to RMSProp will be a matter of just changing the optimizer class. An optimizer now optimizes on a single tensor/vector of weights.

Layers and activation functions will now be combined. Creating a tanh layer will now entail creating a normal layer, and then a tanh layer. This means that testing new layers will be easier, but will make the creating of a neural network a bit more difficult. Expect convenience methods in future versions.