|
"
Utilisation de la stratégie d'apprentissage dans les réseaux connexionnistes
"
Download (in french): |
Learning in most neural networks is based on a passive acquisition of data, according to a basic " stimulus/response " scheme. Those neural networks meet however difficulties to learn complex tasks with a lot of synaptic weights.
In this paper, we consider that those networks lack the ability to structure themselves and structure what they learn. We introduce the notion of learning strategy, a mechanism to control what to learn and how. We propose that learning should occur so as to reduce the "effort" the network has to accomplish to learn the data. Some learning architectures are proposed to achieve such a progressive learning. The proposed architecture has not been developed further, because it was not precise enough. The interest of this paper is more "historical", introducing for the first time the idea of "learning in time" in our research. |
(( TOPICS
Progressive learning (( LABS Lifia McGill |