This week, we are having a double feature seminar (see http:// www.iro.umontreal.ca/article.php3?id_article=107&lang=en):
Image Classification with Higher-Order Neural Models
and
Deep Learning with Denoising Autoencoders
by James Bergstra and Pascal Vincent, Département d’Informatique et de Recherche Opérationnelle Université de Montréal
Location: Pavillon André Aisenstadt (UdeM), room 3195 Time: March 25th 2008, 10h00
First talk, by James Bergstra: Neural network research in machine learning grew out of theories from computational neuroscience from the 1960s. While the class of affine- sigmoidal feature extractors has been studied extensively since the mid 1980s, the computational neuroscience community has moved on to new models that are qualitatively different and more descriptive, without being substantially more computationally expensive. This paper brings a particular model proposed in (Rust, 2005) for low level neurons in the macaque visual system into a machine-learning context: we evaluate their model as an activation function (feature- extractor) for single-layer neural networks that perform image classification. The function we evaluate is somewhat similar to the higher-order processing units discussed in (Minsky, 1969) and the Sigma-Pi units, but avoids the computational difficulties associated with these models by representing the second-order interaction weights with a low-rank positive semi-definite matrix, and avoids the learning difficulties associated with these models by using a gentler non-linearity than the logistic sigmoid. Remarkably good comparative results are obtained on three image classification tasks including 1.4 % error on MNIST using a single-layer network. These results suggest that a single hidden layer neural network equipped with this neuron model can capture important patterns that escape standard models such as sigmoid neural networks and support vector machines based on gaussian and polynomial kernels.
Second talk, by Pascal Vincent: Previous work has shown that the difficulties in learning deep generative or discriminative models can be overcome by an initial unsupervised learning step that maps inputs to useful intermediate representations. We introduce and motivate a new training principle for unsupervised learning of a representation based on the idea of making the learned representations robust to partial corruption of the input pattern. This approach can be used to train autoencoders, and these denoising autoencoders can be stacked to initialize deep architectures. The algorithm can be motivated from a manifold learning and information theoretic perspective or from a generative model perspective. Comparative experiments clearly show the surprising advantage of corrupting the input of autoencoders on a pattern classification benchmark suite.