Tools for analysing deep rectifier networks
A set of tools for analysing deep fully-connected recitifer networks i.e. networks with ReLU activation. Specific tools include methods for experimental analysis of optimisation, variable disentangling, variable size representation, sparsity, and generalisation.
The ultimate goal is to develop useful mathematical models for deep rectifier networks, a model for supervised learning that is not yet well-understood from a theoretical perspective.