Efficient Processing of Spatio-Temporal Data Streams With Spiking Neural Networks.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Deep Learning

Background:

• Spiking neural networks (SNNs) offer high efficiency for neuromorphic hardware but traditional ANN-to-SNN conversion methods struggle with sequence processing due to differing temporal integration.
• Artificial neural networks (ANNs) and SNNs process information over time differently, posing challenges for direct conversion techniques in sequence-based tasks.

Purpose of the Study:

• To develop a novel method for training ANNs that optimizes them for conversion into highly accurate SNNs for sequence processing.
• To enable SNNs to fully exploit the advantages of parallel neuromorphic hardware for temporal data.

Main Methods:

• Modified ANN training prior to conversion to align ANN rollout delays with SNN propagation delays.
• Utilized the streaming rollouts framework for fully parallel ANN execution and temporal integration.
• Incorporated spatio-temporal shortcut connections to improve network responses over time.

Main Results:

• Achieved state-of-the-art accuracy on multiple event-based benchmark datasets (N-MNIST, CIFAR10-DVS, N-CARS, DvsGesture).
• Demonstrated low-latency network responses that improve with increased input sequence processing.
• Converted SNNs showed consistent energy efficiency improvements over their corresponding ANNs.

Conclusions:

• The proposed method successfully bridges the gap between ANN training and SNN conversion for sequence processing tasks.
• This approach unlocks the potential of SNNs for efficient and accurate temporal data analysis on neuromorphic hardware.
• The findings pave the way for more energy-efficient AI systems in real-time applications.