Analysis and pruning of temporally dynamic neural networks

An algorithm for pruning temporally dynamic neural networks is suggested. A set of "importance and similarity measures" are defined for both links and nodes of the network, which are computed recursively from output to input. Pruning and analysis of the network can be performed based on these importance and similarity measures. The suggested algorithm is tested on several networks including a "multi-rate temporal flow model" which is trained for a speaker independent phoneme recognition task.<
>