Work

After having done my PhD at the LISA lab with Yoshua Bengio on neural networks, I worked on probabilistic models of images with the annoyingly good John Winn in the MLP group at Microsoft Research Cambridge.

My interests mostly lie in the fields on deep architectures, semi-supervised learning and online learning.

News

• 12 Mar 2012 - There was a bug in the Matlab code for LCA. It should now be corrected. You can download the files here.


• 28 Feb 2012 - After years of using it, I finally released my Matlab code to train neural networks. It is meant to be fast, but does not yet use any fancy optimization method (just plain old stochastic gradient descent). Download it here.


• 13 Sep 2011 - Our paper Convergence Rates of Inexact Proximal-Gradient Methods for Convex Optimization is now available on arXiv. We obtain bounds on the convergence rate of proximal-gradient methods in the presence of noisy gradients and when the proximal problem is only approximately solved.


• 13 Sep 2011 - Our technical report on Local Component Analysis is now available on arXiv. It is a metric learning algorithm which tries to make the data locally isotropic. It is a useful preprocessing tool for all algorithms assuming local isotropy, like spectral clustering. You can try it using the code available here.


• 17 Mar 2011 - Slides from my SMILE seminar on Deep Belief Networks.

Stuff

• Download my resume (updated on 16/11/11)
• Email: nicolas@le-roux.name