Related Experiment Videos
Predicting the behaviour of G-RAM networks
Geoffrey G Lockwood1, Igor Aleksander
1Department of Anaesthesia, Imperial College School of Medicine, London, UK. g.lockwood@ic.ac.uk
Summary
Generalising Random Access Memory (G-RAM) neurons offer variable tolerance to pattern deviations. Reducing the generalization parameter enhances network stability against training set variations, improving robustness.
Area of Science:
- Artificial Intelligence
- Computational Neuroscience
- Machine Learning
Background:
- Conventional neuron models have fixed tolerance to input variations.
- Digital probabilistic neural networks, also known as n-tuple or weightless systems, are gaining attention.
- Generalising Random Access Memory (G-RAM) neurons introduce variable tolerance to training pattern deviations.
Purpose of the Study:
- To analyze the behavior of recursive G-RAM networks.
- To investigate the impact of training set characteristics on network performance.
- To explore the role of the generalization parameter in G-RAM neuron robustness.
Main Methods:
- Combinatoric analysis of recursive G-RAM network behavior.
- Evaluation of network performance using the 'radius of retrievability' parameter.
- Assessment of training set size and pattern diversity (mean intra-set Hamming distance).
Main Results:
- Optimal network performance is achieved with random training data patterns.
- Increasing training set size or using non-random patterns reduces the radius of retrievability.
- Reducing the G-RAM generalization parameter decreases the radius of retrievability but increases stability against training set changes.
Conclusions:
- Network robustness is enhanced by adjusting the G-RAM generalization parameter.
- The G-RAM generalization parameter offers a novel method to control network stability.
- This study predicts novel behavior regarding system robustness with varying training set properties.