An Interaction between Auxiliary Knowledge and Hidden Nodes on Time to Convergence

摘要

We investigated the effects of providing auxiliary knowledge (or “hints”), of varying the number of hidden nodes, and of providing secondary structure on the performance of feedforward networks with 0, 3, 6, and 9 hidden nodes. Data that permitted the prediction of diabetes from Pima Indian women served as inputs. By systematically adding secondary structure (not obtainable from the original data), we were able to show that convergence time was a function of the number of hidden nodes. The results suggest that neural nets learn “the easy things first” and that providing additional information may impair performance if secondary structure exists in the input data. We propose a model that is consistent with our results and that is also able to account for the common finding that performance on testing sets shows an initial increase followed by a gradual decline to an asymptote.