Storage of Phase-Coded Patterns via STDP in Fully-Connected and Sparse Network: A Study of the Network Capacity

Area of Science:

• Computational Neuroscience
• Artificial Neural Networks

Background:

• Recurrent neural networks (RNNs) are models of neural computation.
• Storing and retrieving information are key functions of neural systems.

Purpose of the Study:

• To investigate the storage and retrieval of phase-coded patterns in recurrent neural networks.
• To analyze network capacity and retrieval dynamics in analog and spiking models.

Main Methods:

• Utilized both analog and integrate-and-fire spiking recurrent neural network models.
• Employed a learning rule based on spike-time-dependent plasticity with asymmetric time windows.
• Examined fully connected and sparse network topologies, including small-world networks.

Main Results:

• Network capacity scales linearly with size in analog models.
• Retrieval state oscillation frequency depends on learning rule asymmetry.
• Sparse networks with a small fraction of long-range connections show amplified capacity, suggesting optimal small-world topology.
• Spiking models also demonstrate successful storage and retrieval of phase-coded patterns.

Conclusions:

• Phase-coded patterns can be stored and retrieved as stable attractors in RNNs.
• Network topology, particularly small-world structures, significantly impacts storage capacity.
• Learning rule asymmetry is a critical factor influencing retrieval dynamics and network performance.