Effect of the Neuron Coding by Gaussian Receptive Fields on Enhancing the Performance of Spiking Neural Network for An Automatic Lipreading System

The artificial neural networks have been generally based on rate coding in the earliest stage of computational neuroscience development. What if all the idea of computational paradigm involving the propagation of continuous data affected straight the enhancing of neural network performance and the main objective becomes how to encode the data for modeling biological behavior. The spiking neural networks (SNN) were founded around this concept where not only the network topology, neuron model and plasticity rule should be defined, but also used the timing of the spike to encode and compute information. In this paper, we proposed an automatic lipreading system for spoken digits based on spike response model (SRM). We experimentally demonstrated the impact of the coding strategy to improve the results by comparing two strategies: Spike time coding and population coding by using Gaussian receptive fields (GRF); which achieved 75% and 83.33% accuracy, respectively, on Tulips1.0 dataset.