Memory-Efficient Synaptic Connectivity for Spike-Timing- Dependent Plasticity.

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering

Background:

• Spike-Timing-Dependent Plasticity (STDP) is a key learning rule for neuromorphic systems, updating synaptic weights based on spike timing.
• Traditional STDP requires access to both forward and reverse synaptic connectivity for weight updates, posing challenges for large-scale systems.
• Memory efficiency and access costs are critical for implementing STDP in large-scale neuromorphic hardware.

Purpose of the Study:

• To investigate the impact of different synaptic connectivity storage arrangements on weight updates in STDP.
• To analyze memory efficiency, storage, and access costs for various connectivity densities.
• To propose and validate an alternative STDP formulation for efficient weight access in sparse networks.

Main Methods:

• Analysis of memory efficiency for synaptic connectivity storage, from dense crossbar arrays to sparse pointer-based lookups.
• Comparison of storage and access costs across different memory arrangements.
• Development and functional validation of a delayed causal update mechanism for STDP.

Main Results:

• Different synaptic connectivity arrangements exhibit varying memory efficiencies and access costs, with trade-offs based on network density and application needs.
• The proposed delayed causal update mechanism for STDP requires only forward connectivity lookup.
• The delayed causal STDP formulation is functionally equivalent to the original, offering substantial savings in storage and access costs for sparse networks.

Conclusions:

• Optimizing synaptic connectivity storage is crucial for efficient STDP implementation in large-scale neuromorphic systems.
• The delayed causal STDP formulation provides a memory-efficient alternative, particularly beneficial for sparse networks common in computational neuroscience.
• This work enables more scalable and efficient neuromorphic learning through optimized STDP implementation.