Abstract:
We study networks of spiking neurons that use the timing of pulses to encode information. Nonlinear interactions model the spatial groupings of synapses on the dendrites and describe the computations performed at local branches. We analyze the question of how many examples these networks must receive during learning to be able to generalize well. Bounds for this sample complexity of learning are derived in terms of the pseudo-dimension. In particular, we obtain almost linear and quadratic upper bounds in terms of the number of adjustable parameters for depth-restricted and general feedforward architectures, respectively. These bounds are also shown to be asymptotically tight for networks that satisfy realistic constraints.