A neural network based on sequence learning: Application to spoken digits recognition

A wide range of real-world problems involves dynamic aspects. Accordingly, some neural models have been proposed that try to incorporate the temporal dimension. Our aim is to design a generic temporal neural model which can learn and schedule spatio-temporal masks. A self-organized mechanism can also be used to preserve the spatio-temporal topology. We define spatio-temporal maps which involve sequences corresponding to patterns at different levels of integration. A simple basic temporal unit is introduced and can be viewed as an evolutive unit to test different algorithms. We first verify the existence of characteristic sequences in the speech signal, then we apply our sequence model to spoken digits recognition. The genericity of our model, its simplicity and its facility to evolve will allow us to use it for other temporal problems.