On-Chip Training Spiking Neural Networks Using Approximated Backpropagation With Analog Synaptic Devices

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Spiking neural networks (SNNs) offer low power and parallel processing potential, mimicking biological systems.
• Supervised training of SNNs for hardware implementation remains a challenge.
• Gated Schottky diodes (GSDs) present a promising synaptic device with saturated current characteristics.

Purpose of the Study:

• To propose an efficient on-chip training scheme for hardware SNNs.
• To approximate the backpropagation algorithm for SNN hardware.
• To evaluate the performance of SNNs trained with the proposed scheme using GSDs.

Main Methods:

• Developed an on-chip training scheme approximating backpropagation for SNNs.
• Utilized gated Schottky diodes (GSDs) as synaptic devices in the SNN hardware.
• Trained and validated the SNN system on the MNIST dataset, analyzing network size and time steps.

Main Results:

• Achieved high classification accuracy on MNIST: 97.83% (1 hidden layer) and 98.44% (4 hidden layers).
• Demonstrated SNN accuracy close to conventional ANNs by leveraging neuronal stochasticity.
• Evaluated the impact of GSD non-linearity, asymmetry, and device variations on SNN performance.

Conclusions:

• The proposed on-chip training scheme is effective for hardware SNNs using GSDs.
• The SNN system demonstrates competitive accuracy and potential for efficient neuromorphic computing.
• Further research can explore device variations and non-linearities for optimized SNN performance.