PATTERN-RECOGNITION WITH THE DECOMPOSED MULTILAYER PERCEPTRON

The Letter reports the benefits of decomposing the multilayer perceptr on (MLP) for pattern recognition tasks. Suppose there are N classes, t hen instead of employing 1 MLP with N outputs, N MLPs are used, each w ith a single output. In practice, this allows fewer hidden units to be used than would be employed in the single MLP. Furthermore, it is fou nd that decomposing the problem in this way allows convergence in fewe r iterations, and it becomes straightforward to distribute the trainin g over as many workstations as there are pattern classes. The speedup is then linear in the number of pattern classes, assuming there arc as many processors as classes. If there are more classes than processors , then the speedup is linear in the number of processors. It is shown that on a difficult hand-written OCR problem, the results obtained wit h the decomposed MLP are slightly superior than those for the conventi onal MLP, and obtained in a fraction of the time.