Memory-efficient neurons and synapses for spike-timing-dependent-plasticity in large-scale spiking networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Simulating large-scale spiking neural networks (SNNs) faces efficiency challenges due to frequent memory access, particularly during synaptic plasticity.
• Synaptic plasticity rules significantly contribute to memory access overhead, limiting simulation performance.

Purpose of the Study:

• To propose novel neuron models and memory access strategies for efficient simulation of SNNs with synaptic plasticity.
• To reduce memory access latency and overhead in large-scale SNN simulations.

Main Methods:

• Developed simplified neuron models with three state variables to enforce neuronal dynamics.
• Implemented memory retrieval focused on postsynaptic variables for contiguous storage and burst mode operations.
• Analyzed memory access patterns compared to naive approaches.

Main Results:

• The proposed method significantly reduces average memory accesses compared to naive methods.
• Achieved contiguous memory storage and leveraged burst mode operations for reduced access overhead.
• Demonstrated the ability to implement different plasticity rules, resulting in varied synaptic weight distributions.

Conclusions:

• The proposed strategy effectively speeds up memory transactions and reduces latency in SNN simulations.
• This approach maintains a small memory footprint while enhancing simulation efficiency.
• The method offers a viable solution for computationally intensive SNN simulations involving synaptic plasticity.