TRAINING A LINEAR MACHINE ON MISLABELED PATTERNS.
Abstract
A linear machine is a multicategory classifier that can be trained to classify a pattern from a set of linearly separable patterns into one of a finite number of categories. It is trained on a representative set of patterns, each pattern having been labeled with a desired category number. If the patterns so labeled are linearly separable, then training with any of several error-correction procedures will yield error-free performance after a finite number of corrections. This report is concerned with the behavior of a linear machine when the training patterns are not linearly separable. Particular attention is devoted to the nonseparable problem arising when the training patterns are randomly mislabeled. Training is viewed as a Markov process, and the behavior of the time-average linear machine is analyzed. It is shown that the time-average linear machine converges to a limiting classifier with probability one, and that this limiting classifier is optimal if the pattern vectors are orthogonal. When the pattern vectors are not orthogonal, the performance may not be optimum; however, experimental results show that averaging usually improves performance of a linear machine quite significantly over that obtained by error-correction alone. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1966
- Accession Number
- AD0634484
Entities
People
- R. O. Duda
Organizations
- SRI International