The Unsupervised and Transfer Learning Challenge is now officially over ! The LISA lab performed very well, winning top honours in Phase 2, which consisted in unsupervised learning of robust features that would transfer to a new distribution with new classes. Full contest results can be found here. Unsurprisingly, our strategy relied heavily on deep learning methods, notably the recent Contractive Auto-Encoder and Spike & Slab RBM.
A JMLR submission detailing our methodology is currently under review for publication in the fall. We will post an update on deeplearning.net when it is made available.