Classification by EM-Trained Dynamic Artificial Neural Nets Based on Hidden Perceptrons,
Abstract
We propose to classify points in R d by functions related to two-layer (a single hidden layer) feedforward artificial neural nets (ANNs). These functions, dubbed dynamic ANNs (DANNs), arise in a rather natural way from probabilistic and also statistical considerations. We treat the binary classification problem and outline an approach to the n-ary classification problem. There are two key ideas. The probabilistic idea is that DANNs are conditional probabilities certain mixture models. The statistical idea is that these models, and hence the DANNs defined by them, are conveniently trainable by an expectation - maximization (EM) algorithm.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1992
- Accession Number
- ADP007152
Entities
People
- Arthur Nadas
Organizations
- IBM Thomas J. Watson Research Center