Workshop on Autoencoders/Autoassociators

NIPS 97 Workshop:
Advances in Autoencoder/Autoassociator-Based Computations

Date: Friday December 5, 1997

Location: Breckenridge, Colorado

Organizers: Nathalie Japkowicz (, Mark A. Gluck ( and Stephen J. Hanson (

Workshop Description: Autoencoders/Autoassociators have had a troubled history. At first believed to have great potential for image compression and speech processing, they were subsequently shown not to outperform Principal Component Analysis (PCA), a linear dimensionality-reduction method, even in the presence of nonlinearities in their hidden layer. Nonetheless, because of their intriguing nature, their study was pursued and it was shown that under certain circumstances, they are capable of performing various types of nonlinear dimensionality-reduction tasks. More recently, they were also shown to be very competent in learning algorithm's reliability estimation, novelty detection, cognitive modeling of the hippocampus and of the natural language grammar acquisition process. Furthermore, they appear promising for time-series analyses when used in recursive mode. Despite their various successes, however, autoencoders/autoassociators have had a difficult time re-establishing themselves fully and the extent of their capabilities remains controversial. The purpose of this workshop is to attempt to revive this extremely powerful, yet conceptually simple device in order to generate a more elaborate understanding of their inner-workings and to explore the various practical and cognitive applications for which they can be useful. More specifically, we are hoping to bring together theoretical and experimental researchers from both engineering and cognitive science backgrounds, who have studied autoencoders/autoassociators, used them in various ways, or designed and studied novel autoencoder/autoassociator-based architectures of interest. We believe that by enabling the sharing of ideas and experiences, this forum will help understand autoencoders/autoassociators better and generate new research directions. In addition, autoencoder/autoassociator-based systems will be compared to other types of autoassociative schemes and their strengths and weaknesses evaluated against them.

Confirmed List of Speakers (with links to talk abstracts and papers, in some cases):