Max-Planck-Institut für Informatik
max planck institut
mpii logo Minerva of the Max Planck Society

MPI-INF or MPI-SWS or Local Campus Event Calendar

<< Previous Entry Next Entry >> New Event Entry Edit this Entry Login to DB (to update, delete)
What and Who
Title:Learning representations with autoencoders: a brief overview and a few novel perspectives
Speaker:Pascal Vincent
coming from:Département d'Informatique et Recherche Opérationnelle Université de Montréal
Speakers Bio:
Event Type:Talk
Visibility:D1, D2, D3, D4, D5, RG1, SWS, MMCI
We use this to send out email in the morning.
Level:Public Audience
Date, Time and Location
Date:Monday, 17 November 2014
Duration:45 Minutes
Building:E1 4
Representations are at the heart of adaptive perception and the ability to take intelligent decisions. Raw low-level sensory inputs typically come in a very high dimensional, intricate representation wherein meaning and statistical regularities are deeply hidden. In recent years, research on novel principles, approaches and algorithms for the unsupervised discovery and learning of good representations -- starting from low-level ones as input -- has been a key enabler of progress in machine learning. Techniques this research inspired are now routinely used to train large deep networks that achieve record-breaking performance on supervised tasks of industrial relevance.

Among the multiple paradigms for unsupervised representation learning, my focus will be on autoencoders. I will give a brief overview of several autoencoder variants that I contributed to develop, and explain how they can be related to other paradigms such as probabilistic graphical models and manifold modeling. This presentation should convey a sense of what was learned, how this line of research matured, and how my vision of what representation learning should consist of changed along the way. I will also present novel perspectives, such as decoderless autoencoders, as well as other research directions and open questions that I think are worth investigating further, in our quest for the ability to autonomously learn truly better, more meaningful, representations.

Name(s):Connie Balzert
Phone:0681 9325-2000
Video Broadcast
Video Broadcast:NoTo Location:
Tags, Category, Keywords and additional notes
Attachments, File(s):
  • Connie Balzert, 11/12/2014 03:16 PM -- Created document.