Related Experiment Videos
Attraction basin of bidirectional association memories
1Department of Computer Science and Engineering, Chinese University of Hong Kong, Shatin, N.T., Hong Kong. csleung@cs.uow.edu.au
International Journal of Neural Systems
|December 1, 1996
Summary
This study analyzes the attraction basin of bidirectional associative memory (BAM) models. We derive statistical dynamics to estimate BAM memory capacity and error bounds for improved recalling performance.
Area of Science:
- Artificial Intelligence
- Computational Neuroscience
Background:
- Bidirectional associative memory (BAM) models are two-layer heteroassociators storing bipolar pairs.
- Conventional energy approaches are insufficient for analyzing BAM attraction basins.
Purpose of the Study:
- Investigate the attraction basin of the bidirectional associative memory (BAM) model.
- Address limitations of conventional energy approaches in BAM analysis.
Main Methods:
- Rigorously derive the statistical dynamics of the BAM model.
- Analyze how error bounds change during recalling for arbitrary initial error patterns.
Main Results:
- The derived dynamics allow estimation of worst-case error attraction basins.
- Memory capacity is determined as alpha rn, dependent on neuron layer ratios.
- Error bounds in retrieved pairs are quantified for a given number of stored pairs.
Conclusions:
- The statistical dynamics provide a method to estimate BAM attraction basins and memory capacity.
- A lower bound for the attraction basin in worst-case errors is established for specific conditions (r=1).