An unsupervised training rule for dynamic information processing
Siamack Haghighi, Lex A. Akers · 2002
The authors have derived an unsupervised training rule for multilayered neural networks. The adaptive weight values are selected to maximize the fraction of a node output variance due to correlations among inputs. The processing units are not required to be fully connected. Applications of this training rule in image processing, including edge extraction, picture segmentation, multirepresentation, and motion detection, are presented.>