Simplelearn: a neural network library for machine learning research

This is a pre-release (v0.1).

Simplelearn is a machine learning library that allows you to define complex models using simple parts. It it meant for those who want to quickly define new models and training algorithms. The code is written to be easily readable, understood, and extended.

All models in Simplelearn are directed acyclic graphs of function objects. It is therefore easy to design models that are more complicated than a simple stack of layers. Thanks to the Theano library, these models automatically compile to C/CUDA code, and are automatically differentiable.

It is easy to optimize loss functions and other outputs with respect to arbitrary variables, be they model parameters (training), input variables (inference), or both. Optimizing w.r.t. inputs is useful for visualizing deep features, or searching for pathological inputs that increase the loss function.

Currently Simplelearn only supports differentiable feed-forward models (no RBMs).