Etna | Non-differentiable Neural Networks (WIP)

Design Goals

The main gloal of Etna is to provide fast and thread-safe implementations of non-different deep learning components including:

• Integer-valued layers
• Combinatorial components
• Discrete activation functions
• Stochastic components

Non-differentiable deep learning components are not so famous since such models can not be optimized using gradient-based algorithms like SGD. But don't worry! Etna components can be optimized using powerful gradient-free optimizers available in Zagros.