Topology optimization of random memristors for input-aware dynamic SNN.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Current machine learning models like GPT-4 and SORA lack the efficiency and adaptability of human brains.
• Key limitations include differences in signal representation, optimization, runtime reconfigurability, and hardware architecture.
• The von Neumann bottleneck remains a significant challenge in conventional computing architectures.

Purpose of the Study:

• To introduce a novel brain-inspired computing system, PRIME (pruning optimization for input-aware dynamic memristive spiking neural network).
• To emulate the brain's spiking mechanisms and structural plasticity using memristive spiking neural networks.
• To enhance energy efficiency and reduce computational load in artificial intelligence hardware.

Main Methods:

• Utilizing spiking neurons to mimic biological neural signaling.
• Optimizing the network topology of random memristive spiking neural networks (SNNs) inspired by structural plasticity.
• Implementing an input-aware early-stop policy to reduce processing latency.
• Leveraging memristive in-memory computing to overcome the von Neumann bottleneck.

Main Results:

• PRIME achieved comparable classification accuracy and inception scores to software baselines on a 40-nm, 256-K memristor-based macro.
• Demonstrated significant energy efficiency improvements of 37.8× and 62.5×.
• Reduced computational loads by 77% and 12.5% with minimal performance degradation.
• Exhibited robustness to stochastic memristor noise.

Conclusions:

• PRIME offers a viable pathway towards highly efficient, brain-inspired neuromorphic computing.
• The system effectively mitigates challenges associated with memristor programming stochasticity and the von Neumann bottleneck.
• PRIME represents a significant advancement in developing hardware that mimics the efficiency and adaptability of the human brain.