Pattern recognition with spiking neurons: performance enhancement based on a statistical analysis

PCNN (pulse coupled neural networks) and more generally spiking-neuron models seem to meet the real-time and robustness constraints necessary in on-board pattern recognition applications. However, efficient learning algorithms are still lacking for such networks. We consider a feedforward network of spiking neurons. The weights and biases are obtained after a simple transformation of those learned with standard backpropagation on a static (standard) neural network. We discuss the conditions under which this transformation gives good recognition rates, in the case of handwritten digit recognition