STATISTICAL PATTERN RECOGNITION REVISITED

The name “Pattern Recognition” was initially coined for a discipline, the history of which dates back to the mid-1950s. One of its basic objectives has been to categorize vectorial stochastic data into discrete classes according to simple metric relationships. The first approaches were based on either the adaptive “delta rule”, which is essentially a Robbins-Monro stochastic approximation procedure, or decision-theoretic classification methods using the well-known Bayes probability formulas. Combined with other classical approaches to describe the class density functions such as the Parzen windows, the problem was thought to be solved completely long time ago. In principle it may be so, but in practice, computational aspects such as training and recognition times must be taken into account. The Neural Network hardware would also favor more straightforward operations amenable to analog computing. It has now turned out that certain new self-organizing approaches, here grouped under the title Learning Vector Quantization (LVQ), can combine theoretically near-optimal accuracy with very high speed in learning and recognition; moreover, their implementation by analog VLSI is straightforward. This presentation reports recent experiments, mainly relating to automatic recognition of speech, which demonstrate that there still exist fresh views and effective new solutions to traditional problems.