A neuromorphic implementation of multiple spike-timing synaptic plasticity rules for large-scale neural networks

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence Hardware

Background:

• Synaptic plasticity is crucial for neural network learning and memory.
• Existing neuromorphic systems often lack flexibility in implementing diverse plasticity rules.
• Efficient hardware implementations are needed to support large-scale, complex neural network models.

Purpose of the Study:

• To present a novel neuromorphic implementation supporting multiple synaptic plasticity learning rules.
• To demonstrate a flexible and scalable architecture for advanced neural network development.
• To validate the performance of the proposed digital and mixed-signal implementations.

Main Methods:

• Developed a generic synaptic plasticity adaptor array, separate from neurons.
• Utilized a dynamic-assignment time-multiplexing approach for efficient resource utilization.
• Implemented both fully digital and mixed-signal circuit designs.
• Supported up to 64 million (2^26) synaptic plasticity elements.

Main Results:

• Successfully implemented and validated both Spike Timing Dependent Plasticity (STDP) and Spike Timing Dependent Delay Plasticity (STDDP).
• Demonstrated the flexibility of the adaptor array, allowing multiple plasticity rules without network reconfiguration.
• Measurement results confirmed the functional capabilities of the proposed circuits.
• Showcased the potential for scaling to 64 billion (2^36) synaptic adaptors on FPGAs.

Conclusions:

• The proposed neuromorphic implementation offers a flexible and scalable platform for diverse synaptic plasticity rules.
• The generic adaptor array architecture significantly enhances the adaptability of large-scale neural networks.
• The technology is practical for scaling to extremely large neural network sizes on current hardware.