Catalogue of Artificial Intelligence Techniques
Time-delay Neural Networks
Keywords: back-propagation, neural networks
Categories: Neural Networks
Author(s): Steve Renals
A limitation of feed-forward networks is their inability to model time-dependent data. A proposed solution to this problem is the addition of several, time-delayed connections between a pair of units. In this way time-dependencies may be modelled. Training may be performed using the usual back-propagation of error algorithm (see Multi-layer Perceptorns). The number of free parameters is reduced by constraining a set of connections differing only in delay to have the same weight--these constraints also impart a translation invariance into the system.
- Hirai, A. and Waibel, A., Phoneme-based word recognition by neural network: a step toward large vocabulary recognition
IJCNN-90: proceedings of the international joint conference on neural networks
3 (1990), Ann Arbor, MI: IEEE Neural Networks Council, 671--676.