Are binary synapses superior to graded weight representations in stochastic attractor networks?

Area of Science:

• Neuroscience
• Computational Neuroscience
• Theoretical Neuroscience

Background:

• Synaptic plasticity underlies learning and memory.
• Debate exists on whether synaptic efficacy is graded or binary.
• Binary weights are hypothesized to be more noise-resistant than graded weights.

Purpose of the Study:

• Compare retrieval performance of binary vs. graded synaptic weight representations.
• Investigate models with multiple discrete weight states.
• Determine optimal dilution threshold for binary weight representations.

Main Methods:

• Numerical simulations of stochastic attractor networks.
• Comparison of model retrieval performance with different weight representations.
• Analysis of load capacity and noise susceptibility.

Main Results:

• Binary weight representations are not less susceptible to noise than graded representations in stochastic attractor models.
• Increasing the number of discrete weight states rapidly approaches the load capacity of graded weights.
• Optimal dilution threshold for binary weights under stochastic conditions is achieved when ~50% of the smallest weights are set to zero.

Conclusions:

• The assumption of superior noise resistance for binary synaptic weights in stochastic attractor models is challenged.
• Discrete weight states offer a viable alternative, with performance approaching graded weights as states increase.
• A specific dilution strategy can optimize binary weight representations in noisy neural network models.