Fast agreement-driven device-calibrated local learning paradigms for spiking neural networks

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Classical Spike-Timing-Dependent Plasticity (STDP) relies on precise spike timing, limiting computational efficiency.
• Spiking Neural Networks (SNNs) offer potential for energy-efficient AI but require effective learning rules.
• Biologically inspired learning rules are crucial for advancing neuromorphic computing.

Purpose of the Study:

• Introduce novel synaptic learning rules: Spike Agreement Dependent Plasticity (SADP) and Spike Correlation Dependent Plasticity (SCDP).
• Evaluate the efficiency and effectiveness of SADP and SCDP compared to classical STDP.
• Explore the potential of agreement-driven plasticity for neuromorphic systems.

Main Methods:

• Developed SADP and SCDP, focusing on spike train agreement and correlation, not precise timing.
• Implemented SADP and SCDP with linear-time complexity for efficient hardware realization.
• Validated SADP and SCDP on MNIST and Fashion-MNIST datasets using single-layer networks.

Main Results:

• SADP and SCDP demonstrated significantly faster feature learning than STDP.
• Achieved strong downstream classification accuracy with an order of magnitude less training time.
• Showcased efficient hardware implementation capabilities via event-driven, bitwise logic.

Conclusions:

• Agreement-driven plasticity provides a fast, local, and biologically plausible learning framework.
• SADP and SCDP are viable unsupervised learning mechanisms for neuromorphic systems.
• The framework offers potential extensions to supervised learning paradigms for enhanced AI capabilities.