Pruning for Hardware-Based Deep Spiking Neural Networks Using Gated Schottky Diode as Synaptic Devices.

Area of Science:

• Artificial Intelligence
• Neuromorphic Engineering
• Hardware Acceleration

Background:

• Deep neural networks (DNNs) offer state-of-the-art performance but suffer from high computational costs, hindering real-time applications.
• Spiking neural networks (SNNs) present a computationally efficient alternative to DNNs for energy-constrained environments.
• Gated Schottky diodes offer a promising solution for efficient synaptic device implementation in neuromorphic hardware.

Purpose of the Study:

• To propose and evaluate a hardware implementation of SNNs utilizing gated Schottky diodes as synaptic devices.
• To investigate the application of L1 regularization for efficient weight pruning in SNNs, eliminating the need for retraining.
• To demonstrate the energy efficiency and classification accuracy of the compressed, hardware-based SNN.

Main Methods:

• Hardware implementation of SNNs using gated Schottky diodes as synaptic elements.
• Application of L1 regularization for direct connection pruning of the SNN weights.
• Evaluation of classification accuracy and energy efficiency of the developed hardware SNN.

Main Results:

• The hardware-based SNN achieved a classification accuracy of 97.85%.
• The proposed method, using L1 regularization, successfully pruned network connections without requiring a retraining phase.
• The compressed SNN demonstrated significant energy efficiency compared to traditional DNNs.

Conclusions:

• Hardware implementation of SNNs with gated Schottky diodes offers an energy-efficient solution for real-time machine learning tasks.
• L1 regularization is an effective technique for compressing SNNs, enabling efficient hardware deployment.
• The developed SNN achieves performance comparable to DNNs while significantly reducing computational cost and energy consumption.