Talks

This page links the slides of my most significant lectures. All the slides are encoded in DjVu format. You might have to install the DjVu browser plugin to see them.

This lecture is co-authored with Yann Le Cun. It took place at the 1996 NIPS Workshop Tricks of the Trade organized by Klaus-Robert Müller and Genevieve Orr.

• See the slides (djvu 199KB) (pdf 2.7MB).
• See the corresponding book chapter.

This lecture describe Graph Transformer Networks It took place at the 2001 ICML workshop Machine Learning for Spatial and Temporal Data organized by Tom Dietterich. Graph Transformer Networks are one of the most powerful and successful method for learning sequential data. About 10% to 20% of the checks written in the U.S. since 1996 have been processed by a Graph Transformer Network. Graph Transformer Networks are related to Conditional Random Fields but have variable geometry and non-linear energies.

• See the slides (djvu 234KB) (pdf 2.4MB).
• See a short paper or a long paper.

DjVu is a document compression system that allows the distribution of scanned documents on the web. DjVu files are very compact. A typical 300dpi bitonal page takes 10-15KB. A typical 300dpi color page takes 40-60KB. The presentation discusses the main technical innovations that made DjVu possible.

• See the slides (djvu 417KB) (pdf 4.9MB low quality).
• See the DjVu research page and the DjVu software page.

This four part lecture was given at the Machine Learning Summer School held in Tübingen in 2003 organized by Olivier Bousquet, Bernhard Schölkopf and Ulrike von Luxburg. The lecture discusses Stochastic Approximations and in particular Stochastic Gradient Descent applied to online learning algorithms.

• See part 1: Framework (djvu 157KB) (pdf 2.1MB).
• See part 2" Cookbook (djvu 459KB) (pdf 2.7MB).
• See part 3: Convergence and Martingales (djvu 38KB) (pdf 1.3MB).
• See part 4: Optimal online algorithms (djvu 55KB) (pdf 1.4MB).

• See the corresponding book chapter.