The memory-based reasoning hypothesis is based on the following assumption: There is no general way to search memory for the best match without examining every element of memory. This paper describes an experimental memory-based reasoning system for pronouncing English words. The system has been implemented on a parallel architecture, specifically the Connection Machine. The algorithm seeks to make decisions by “remembering” similar circumstances in the past. This is done by (1) counting combinations of features, (2) using these counts to produce a metric, (3) using the metric to find the dissimilarity between the current problem and every item in the memory, and (4) retrieving the best matches. Von Neumann machines do not support such a task well. The ideal machine for memory-based reasoning would have a parallel, highly interconnected, fine-grained SIMD architecture.

More important than the details of implementation and the experimental results are the conclusions and prospects derived from such an approach. Emphasis is put on the differences between memory-based and rule-based reasoning. The paper excellently fulfills its basic purpose--to explore a possible reasoning framework that can play an important role in building truly intelligent systems.