Attractor neural networks with double well synapses.

Area of Science:

• Computational Neuroscience
• Neuroscience
• Artificial Intelligence

Background:

• Memory storage is thought to rely on activity-dependent synaptic modifications.
• Synaptic efficacy has been modeled as either continuous or discrete, with recent findings suggesting an intermediate scenario.

Purpose of the Study:

• To explore a novel model of synaptic plasticity with continuous variables in a multi-well potential.
• To investigate how this model interpolates between discrete and continuous synaptic models.
• To analyze the impact of synaptic potential shape on information storage capacity and robustness.

Main Methods:

• Developed an analytical and numerical model of neural networks with multi-well synaptic potentials.
• Analyzed the storage capacity scaling with network size.
• Investigated the effect of potential well depth and noise on memory storage.

Main Results:

• The model successfully interpolates between discrete (deep potential) and continuous (single quadratic potential) synapse models.
• Networks with double-well synapses show a power-law dependence of storage capacity on network size, unlike the logarithmic dependence in single-well models.
• Deeper potential wells enhance information storage robustness against noise.

Conclusions:

• Synaptic efficacy described by continuous variables in multi-well potentials offers a more nuanced understanding of memory storage.
• This intermediate model provides a framework for improved information storage in artificial neural networks.
• The findings suggest potential advantages for robust memory formation in biological and artificial systems.