When Are k-Nearest Neighbor and Back Propagation Accurate for Feasible Sized Sets of Examples?

We first review in pedagogical fashion previous results which gave lower and upper bounds on the number of examples needed for training feedforward neural networks when valid generalization is desired. Experimental tests of generalization versus number of examples are then presented for random target networks and examples drawn from a uniform distribution. The experimental results are roughly consistent with the following heuristic: if a database of M examples is loaded onto a W weight net (for M≫W), one expects to make a fraction ɛ=W/M errors in classifying future examples drawn from the same distribution. This is consistent with our previous bounds, but if reliable strengthens them in that: (1) the bounds had large numerical constants and log factors, all of which are set equal one in the heuristic, (2) previous lower bounds on number of examples needed were valid only in a distribution independent context, whereas the experiments were conducted for a uniform distribution, and (3) the previous lower bound was valid for nets with one hidden layer only. These experiments also seem to indicate that networks with two hidden layers have Vapnik-Chervonenkis dimension roughly equal to their total number of weights.