Authors:
Mohammad Saifullah
;
Rita Kovordanyi
and
Chandan Roy
Affiliation:
Linköping University, Sweden
Keyword(s):
Generalization, Image Processing, Bidirectional Hierarchical Neural Networks, Hebbian Learning, Feature Extraction, Object Recognition.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Biomedical Engineering
;
Biomedical Signal Processing
;
Computational Intelligence
;
Computer Vision, Visualization and Computer Graphics
;
Data Manipulation
;
Feature Extraction
;
Features Extraction
;
Health Engineering and Technology Applications
;
Human-Computer Interaction
;
Image and Video Analysis
;
Image Formation and Preprocessing
;
Implementation of Image and Video Processing Systems
;
Informatics in Control, Automation and Robotics
;
Methodologies and Methods
;
Neural Networks
;
Neurocomputing
;
Neurotechnology, Electronics and Informatics
;
Pattern Recognition
;
Physiological Computing Systems
;
Sensor Networks
;
Signal Processing
;
Signal Processing, Sensors, Systems Modeling and Control
;
Soft Computing
;
Theory and Methods
Abstract:
Visual pattern recognition is a complex problem, and it has proven difficult to achieve satisfactorily in standard three-layer feed-forward artificial neural networks. For this reason, an increasing number of researchers are using networks whose architecture resembles the human visual system. These biologically-based networks are bidirectionally connected, use receptive fields, and have a hierarchical structure, with the input layer being the largest layer, and consecutive layers getting increasingly smaller. These networks are large and complex, and therefore run a risk of getting overfitted during learning, especially if small training sets are used, and if the input patterns are noisy. Many data sets, such as, for example, handwritten characters, are intrinsically noisy. The problem of overfitting is aggravated by the tendency of error-driven learning in large networks to treat all variations in the noisy input as significant. However, there is one way to balance off this tendency
to overfit, and that is to use a mixture of learning algorithms. In this study, we ran systematic tests on handwritten character recognition, where we compared generalization performance using a mixture of Hebbian learning and error-driven learning with generalization performance using pure error-driven learning. Our results indicate that injecting even a small amount of Hebbian learning, 0.01 %, significantly improves the generalization performance of the network.
(More)