Related Experiment Videos
Sparsification from dilute connectivity in a neural network model of memory
1Institute for Theoretical Physics, State University of New York at Stony Brook, 11794-3840, USA. mmaraval@insti.physics. sunysb.edu
Network (Bristol, England)
|June 18, 1999
Summary
This study reveals that simple network models for brain associative memory need to consider activation thresholds. Evaluating network performance requires balancing pattern storage with information preservation, not just capacity.
Area of Science:
- Neuroscience
- Computational Neuroscience
- Artificial Intelligence
Background:
- Hypotheses on brain associative memory often use simple network models.
- Cerebral network's low connectivity imposes constraints not always clear in these models.
Purpose of the Study:
- Investigate the overlooked aspect of activation threshold setting in dilute neural networks.
- Determine optimal criteria for threshold assignment in associative memory models.
- Propose a more comprehensive evaluation of network performance.
Main Methods:
- Analysis of simple, dilute network models.
- Examination of various criteria for optimal activation threshold assignment.
- Derivation of network sparsification and representational ability loss due to dilution.
Main Results:
- Network capacity alone is insufficient to characterize performance quality.
- Dilution leads to sparsification (decreased firing probability).
- Sparisfication results in losses in representational ability.
Conclusions:
- Activation threshold setting is a critical, often overlooked, factor in neural network models of associative memory.
- Network evaluation must account for sparsification and representational losses.
- A trade-off between pattern storage and information preservation is essential for assessing model suitability.