Shape of error surfaces in SpikeProp

In this paper, we discuss the shape of error surfaces, which represent error depending on parameters, in Spiking Neural Networks for SpikeProp. SpikeProp is a learning algorithm that adjusts timing of spikes. The discussion is held in the viewpoint of the difference between analogue computation and digital computation (especially in discrete time). Since the error is defined by timing of spikes, quantization error brought by digital computation changes the shape. We show typical shapes of error surfaces through some experiments. Digital computation bring rough error surfaces, which have many false local minima. These local minima will disturb effective acceleration of learning process by sophisticated algorithms.