Using 3-D surface maps to illustrate neural network performance

A possible means for evaluating the performance of neural networks from a global perspective in parameter-space is suggested. An organized experimental method that identifies network configuration and parameter value choices which are not sensitive to minor variations for a standard training metric is described. Convergence maps are n-dimensional plots which show the ability of a neural network to converge on (learn) a given training metric. The traveling salesman optimization problem is a classic metric for testing energy minimization networks. This metric isa discussed. The technique is illustrated for the network used by J.J Hopfield and D.W. Tank (1985) to solve a traveling salesman problem and with traditional backpropagation as described by R.P Lippmann (1987).<
>