|
"
Les réseaux récurrents supervisés : une revue critique
"
Download (in french): |
This report presents a detailed survey on supervised recurrent neural networks. These networks can be distinguished according to the type of behaviour they learn, their structure, the discrete or continuous formalism and the learning algorithm. The two main learning algorithms are reported, in a didactic and synthetic manner. In particular, variants of Back Propagation Through Time are replaced in a common framework.
Then, several analysis and comparisons of these algorithms are performed, allowing a better understanding of their different features. In particular, an original analysis of the locality in space concept is proposed. All these analysis underscore the difficulties of learning in these networks, leading us to report different variants that have been proposed to overcome these difficulties. |
(( TOPICS
Recurrent nets (( LABS Lifia |