A low cost neuromorphic learning engine based on a high performance supervised SNN learning algorithm.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Computer Science

Background:

• Spiking neural networks (SNNs) offer superior energy and resource efficiency compared to traditional artificial neural networks (ANNs).
• Supervised learning in SNNs is hindered by the non-differentiable nature of spikes and complex computational requirements.
• Designing efficient SNN learning engines faces challenges due to hardware limitations and strict energy budgets.

Purpose of the Study:

• To propose a novel, hardware-efficient back-propagation scheme for SNNs that achieves fast convergence.
• To develop a multiplier-less inference engine and a high-speed training engine for SNNs.
• To demonstrate the effectiveness of the proposed scheme on the MNIST dataset.

Main Methods:

• Introduced a hardware-efficient SNN back-propagation scheme, HaSiST (hard sigmoid SNN training), avoiding complex operations like error normalization.
• Developed a multiplier-less inference engine and a cost-efficient SNN training engine implemented on a Virtex 6 FPGA.
• Evaluated the scheme's performance on the MNIST dataset.

Main Results:

• Achieved approximately 97.5% accuracy on the MNIST dataset using only 158,800 synapses.
• The inference engine operates at 135 MHz, consuming minimal hardware resources per synapse and achieving 9.44 GSOPS.
• The training engine operates at 50 MHz with low resource utilization per synapse.

Conclusions:

• The proposed HaSiST scheme provides an effective and efficient solution for supervised SNN learning.
• The developed hardware engines demonstrate the practical feasibility of deploying SNNs in resource-constrained environments.
• This work contributes to advancing energy-efficient AI through neuromorphic computing.