Proceedings PaperPattern recognition, similarity, neural nets, and optics
|Format||Member Price||Non-Member Price|
The arbitrary nature of similarity and invariance is examined and its implications for pattern recognition and classification are examined. Various measures of similarity are discussed and techniques for achieving invariance under translation rotation contrast and energy are briefly reviewed. We show how both matched filters and neural nets can achieve c!assification of objects into arbitrary classes. For neural nets different kinds of similarity measures can cause patho!ogica! behavior that can be avoided by using a specific normalized kind of similarity measure. Implications for unsupervised learning in certain kinds of neural networks !ike the are discussed.