Algorithm-hardware co-design of neuromorphic networks with dual memory pathways.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) are adept at event-driven processing but struggle with long-term context maintenance under strict energy and memory constraints.
• Existing SNNs face challenges in balancing algorithmic and hardware requirements for sustained context.
• Efficiently managing task-relevant information over extended periods is crucial for advanced SNN applications.

Purpose of the Study:

• To develop an algorithm-hardware co-design for SNNs that enhances long-term context retention.
• To introduce a novel neural network architecture inspired by biological brain organization.
• To improve the energy efficiency and throughput of SNNs for real-time computation.

Main Methods:

• Introduced a dual memory pathway architecture with explicit slow and fast memory pathways inspired by cortical organization.
• Developed a compact, low-dimensional state representation within each layer to modulate spiking dynamics.
• Implemented a near-memory-compute architecture optimizing data flow for sparse-spike and dense-memory pathways.

Main Results:

• Achieved competitive accuracy on long-sequence benchmarks with 40-60% fewer parameters than state-of-the-art SNNs.
• Demonstrated a fourfold increase in throughput compared to existing SNN implementations.
• Showcased a fivefold improvement in energy efficiency through the proposed co-design.

Conclusions:

• Biological principles can guide the development of effective and hardware-efficient algorithmic abstractions for SNNs.
• The proposed dual memory pathway and near-memory-compute architecture offer a scalable framework for real-time neuromorphic computation and learning.
• This approach successfully addresses the challenge of long-timescale context maintenance in SNNs while respecting resource limitations.