Abstract
In this paper, we discuss the influence of a discrete time computation in SpikeProp, which is a kind of training algorithm for spiking neural networks. It adjusts timing of spikes according to the gradient of their error. Digital computation, especially computation in discrete time, causes some quantization errors. We discuss the influence of the quantization error based on the shape of error surfaces, which represent error depending on parameters. Through some experiments, we show learning processes degraded by a digital computation, and a typical shape of error surfaces that cause the degradation. Digital computation brings rough error surfaces, which have many false local minima. These local minima block the effective acceleration brought by sophisticated optimization algorithms.
