Related Experiment Videos
Bistable gradient networks. II. Storage capacity and behavior near saturation
Patrick N McGraw1, Michael Menzinger
1Department of Chemistry, University of Toronto, Toronto, Ontario, Canada M5S 3H6.
Summary
This study numerically investigates bistable gradient networks, finding weak coupling allows higher storage capacity but imperfect pattern retrieval, unlike strong coupling
Area of Science:
- Computational neuroscience
- Artificial neural networks
- Statistical physics
Background:
- Attractor neural networks (ANNs) are models of memory storage.
- The Hopfield network exhibits a
- memory blackout
- phase transition at high storage loads.
- Bistable elements are fundamental units in some neural models.
Purpose of the Study:
- To numerically examine the storage capacity and saturation behavior of bistable gradient networks.
- To compare network behavior under strong versus weak coupling strengths.
- To investigate the trade-off between storage capacity and retrieval accuracy.
Main Methods:
- Numerical simulations of a bistable gradient network.
- Analysis of network behavior near saturation.
- Varying coupling strength as a key parameter.
Main Results:
- Strong coupling shows a first-order
- memory blackout
- phase transition, similar to Hopfield networks.
- Weak coupling avoids this transition, allowing stable patterns at high loads.
- Enhanced storage capacity under weak coupling correlates with imperfect retrieval from corrupted patterns.
Conclusions:
- Bistable gradient networks offer tunable properties for memory storage.
- Weak coupling presents a viable strategy for increased capacity at the cost of fidelity.
- The findings provide insights into designing ANNs with specific storage and retrieval characteristics.