Master Class: Non-local Manifold Learning by Regularized Auto-encoders

SpeakerYoshua Bengio
AffiliationDepartment of Computer Science and Operations Research, University of Montreal
DateTuesday, 22 Oct 2013
Time13:00 - 14:00
LocationGalton Lecture Theatre, 1-19 Torrington Place
Event seriesMaster Class: Yoshua Bengio (21 Oct - 24 Oct 2013)
Description

This talk is about the geometric and probabilistic aspects of regularized auto-encoders. They are shown to be manifold learning algorithms, i.e., exploiting priors about the concentration of probability mass. Whereas most manifold learning algorithms only exploit smoothness to achieve local generalization, we show how regularized auto-encoders are unsupervised learning algorithms that capture the manifold structure and can generalize far from the training examples, predicting the leading tangent vectors of the manifold at any new point, and producing a representation that expands local directions of variation while contracting the others. Whereas they have been mostly used for unsupervised feature learning and supervised fine-tuning, recent work shows probabilistic interpretations allowing to sample from the associated models of the data generating distribution

The video recording of this masterclass is available here:
Oct 22: Non-local Manifold Learning by Regularized Auto-encoders

This event is funded by DeepMind Technologies.

iCalendar csml_id_149.ics