ICML 2011 workshop on unsupervised and transfer learning

Unsupervised and Transfer Learning Challenge: a Deep Learning approach

Gregoire Mesnil1;2
Yann Dauphin1
Xavier Glorot1
Salah Rifai1
Yoshua Bengio1
Ian Goodfellow1
Erick Lavoie1
Xavier Muller1
Guillaume Desjardins1
David Warde-Farley1
Pascal Vincent1
Aaron Courville1
James Bergstra1

1 Dept. IRO, Universite de Montreal. Montreal (QC), H2C 3J7, Canada
2 Dept. LITIS, Universite de Rouen. Mont-Saint-Aignan, 76130, France

Learning good representations from a large set of unlabeled data is a particularly challenging task. Recent work (see Bengio (2009) for a review) shows that training deep architectures is a good way to extract such representations, by extracting and disentangling gradually higher-level factors of variations characterizing the input distribution. In this paper, we describe different kinds of layers we trained for learning representations in the setting of the Unsupervised and Transfer Learning Challenge. The strategy of our team won the final phase of the challenge. It combined different 1-layer unsupervised learning algorithms, adapted to each of the five datasets of the competition. This paper describes that strategy and the particular 1-layer learning algorithms feeding a simple linear classier with a tiny number of labeled training samples (1 to 64 per class). Keywords: Deep Learning, unsupervised learning, transfer learning,