Real-Time Neuromorphic System for Large-Scale Conductance-Based Spiking Neural Networks

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Investigating human intelligence and brain complexity requires advanced computational platforms.
• Simulating large-scale spiking neural networks (LaCSNN) demands high biological realism and network scale.
• Existing neuromorphic systems face challenges in resource efficiency and communication for large-scale simulations.

Purpose of the Study:

• To present a real-time digital neuromorphic system for simulating large-scale conductance-based spiking neural networks (LaCSNN).
• To demonstrate a cost-efficient approach for neuron models and a novel router architecture for efficient communication.
• To enable detailed simulations of complex brain circuits like the cortico-basal ganglia-thalamocortical loop.

Main Methods:

• Development of a real-time digital neuromorphic system using field-programmable gate arrays.
• Implementation of a scalable 3-D network-on-chip (NoC) topology to simulate 1 million neurons.
• Proposal of cost-efficient conductance-based neuron models and a novel router architecture for multinuclei neural networks.

Main Results:

• The system successfully simulated a large-scale cortico-basal ganglia-thalamocortical loop with 1 million neurons.
• Proposed neuron models achieved significant resource reduction (95% less memory, 100% less DSP).
• The novel NoC router architecture effectively managed multiple data flows in the multinuclei network.

Conclusions:

• The developed LaCSNN system offers high biological realism and large network scale for neural simulations.
• The system demonstrates superior performance compared to state-of-the-art approaches for LaCSNN implementation.
• This real-time computational power opens avenues for broad applications in neuroscience and AI.