[Lisa_seminaires] UdeM-McGill-MITACS machine learning seminar Fri March 4th @ 14h00, AA3195

Guillaume Desjardins guillaume.desjardins at gmail.com
Mar 1 Mar 14:23:58 EST 2011


A UdeM-McGill-MITACS machine learning seminar will be held this
Friday, March 4th. The talk given by Hugo Larochelle, will take place
from 14h00-15h00 in the room AA3195 (Université de Montréal). Hope to
see you there !


Title: Two new autoencoders for distribution estimation and guided
representation, Hugo Larochelle (University of Toronto)

Abstract:

In this talk, I'll describe two new autoencoder-like models, developed
for two different problems.

The first is the estimation of distributions of high-dimensional data,
for which the restricted Boltzmann machine (RBM) has been shown to be
a powerful model. However, an RBM typically does not provide a
tractable distribution estimator, since evaluating the probability it
assigns to some given observation requires the computation of the
partition function, which itself is usually intractable. The model
I'll describe circumvents this difficulty by decomposing the joint
distribution of observations into tractable conditional distributions
and modeling each conditionals with a non-linear function similar to
RBM conditionals. This model can also be interpreted as an autoencoder
wired such that its output can be used to assign valid probabilities
to observations. I'll present experiments showing that this new model
outperforms other multivariate binary distribution estimators on
several datasets and performs similarly to a large (but intractable)
RBM.

The second problem is that of guiding an autoencoder toward
representations that are more useful for particular discriminative
tasks. A complementary challenge is finding codes that are explicitly
invariant to irrelevant transformations of the data.  I'll describe
how this can be achieved by combining an autoencoder with a Gaussian
process latent variable model. This enables an autoencoder's
unsupervised representation to both incorporate relevant label
information and ignore irrelevant variations.  I'll describe
experiments on several different datasets which show how both labels
and nuisance variables can provide cues for useful latent
representations.

This is joint work with Iain Murray, Jasper Snoek and Ryan Prescott Adams.


Plus d'informations sur la liste de diffusion Lisa_seminaires