TOMORROW'S UdeM-McGill-MITACS seminar (see http://www.iro.umontreal.ca/article.php3?id_article=107&lang=en) :

Differentiable Sparse Coding

by J. Andrew Bagnell Carnegie Mellon Robotics Institute and Machine Learning Department

Location : Pavillon André-Aisenstadt (UdeM) room 3195 Date and Time : Friday, August 22 2008, 14h00

Sparse approximation is a key technique developed recently in engineering and the sciences which attempts to approximate an input signal, denoted here by X, in terms of a “sparse” combination of fixed bases B. This approach relies on an optimization algorithm to infer the most probable weights \hat{W} to reconstruct input signals, given the model X ≈ f (BW). Priors which produce sparse solutions for W , especially L1 regularization, have gained attention because of their usefulness in ill-posed engineering problems ranging from geology to magnetic resonance imaging, their ability to elucidate certain neuro-biological phenomena, and their ability to condense a high-dimensional input signal into useful features for classification.

Sparse coding – closely connected to Independent Component Analysis as well as certain approaches to matrix factorization – extends sparse approximation by not only performing optimization to compute the best set of weights for a given input signal, but also learning bases B which lead to a compact representation of input signals. Unfortunately, existing sparse coding algorithms that efficiently infer the latent weight vector are difficult to integrate into larger learning architectures. It has been convincingly demonstrated that back-propagation is a crucial tool for tuning an existing generative model’s performance discriminatively to lead to good supervised performance. Similarly, greedy layer-wise strategies to building deep generative models rely upon a back-propagation step to achieve excellent model performance. Existing sparse coding architectures produce a latent representation \hat{W} that is an unstable, discontinuous function of the inputs and bases; an arbitrarily small change in input can lead to the selection of a completely different set of latent weights.

We present a new approach to coding with an efficient, convex inference step based on minimizing KL-divergence. We show this increased stability leads to better semi-supervised classification performance. Additionally, although inferring the latent weights requires an optimization procedure (i.e. is not closed form) we demonstrate that for a large class of Bregman-divergence based priors and loss functions, we may use implicit differentiation to efficiently backpropagate error signals. The sparse coding bases can then be optimized discriminatively leading to outstanding empirical performance and enabling sparse coding to form a part of a larger learning architecture.

Joint work with David M. Bradley