Séminaire ICI : Vincent Gripon
Titre du séminaire et orateur
Neural coding: from error correcting codes to associative memories.
Vincent Gripon, Télécom-Bretagne.
Date et lieu
Mardi 27 novembre 2012, 14h.
ENSEA, Cergy-Pontoise, salle 384.
Associative memories are devices that are able to retrieve previously stored messages given part of their content. Despite the naive ML implementation that would consist in checking successively all stored messages, classical methods are based on neural networks. Those offer a very limited diversity of storage (number of stored messages) as well as a low efficiency (ratio of the amount of bits stored to the amount of bits used).
To considerably increase these two parameters, we propose to use the principles of distributed error correcting codes (turbo, LDPC, ...). We introduce two original codes, and show that when used in associative memories, they allow to achieve nearly-optimal efficiencies. These new devices arose open questions in coding and information theory, as well as a new perspective to better understand the brain memory.