Part of Advances in Neural Information Processing Systems 6 (NIPS 1993)
Mark Plutowski, Garrison Cottrell, Halbert White
We apply active exemplar selection (Plutowski &. White, 1991; 1993) to predicting a chaotic time series. Given a fixed set of ex(cid:173) amples, the method chooses a concise subset for training. Fitting these exemplars results in the entire set being fit as well as de(cid:173) sired. The algorithm incorporates a method for regulating network complexity, automatically adding exempla.rs and hidden units as needed. Fitting examples generated from the Mackey-Glass equa(cid:173) tion with fractal dimension 2.1 to an rmse of 0.01 required about 25 exemplars and 3 to 6 hidden units. The method requires an order of magnitude fewer floating point operations than training on the entire set of examples, is significantly cheaper than two contend(cid:173) ing exemplar selection techniques, and suggests a simpler active selection technique that performs comparably.