This is an extension of the Torch3 Machine Learning library for handling various types of Deep Architectures and modifications to the standard Multi-layer Perceptrons:

• Handles an arbitrary number of fully-connected sigmoidal layers
• Unsupervised learning of MLPs using various reconstruction costs. Greedy layer-wise learning is available as well.
• An implementation of the Stacked Denoising Autoencoders
• A preliminary implementation of collective learning idea, whereby a pair of networks are trained in parallel and are communicating with each other.

If you're new to this page, be sure to check out the following wiki pages:

• InstallationInstructions
• UsageManual
• BackgroundInformation
• RelevantPublications
• BenchmarkDatasets
• FAQ

Math and diagrams describing the algorithms implemented in the code are forthcoming.