Zero-shot temporal resolution domain adaptation for spiking neural networks

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) offer energy efficiency and low latency for temporal data processing on neuromorphic hardware.
• SNN performance degrades significantly when training and deployment data temporal resolutions mismatch, especially without fine-tuning capabilities.
• Existing adaptation methods, like time constant scaling, are often insufficient for resolution mismatches.

Purpose of the Study:

• To develop novel domain adaptation techniques for SNNs to address temporal resolution discrepancies.
• To enable SNN parameter adaptation without requiring retraining on target resolution data.
• To provide a robust solution for deploying SNNs across varying temporal data resolutions.

Main Methods:

• Proposed three novel domain adaptation methods based on a mapping between SNN neuron dynamics and State Space Models (SSMs).
• Applied methods to general neuron models, demonstrating broad applicability.
• Evaluated performance on spatio-temporal datasets: audio keyword spotting (SHD, MSWC) and neuromorphic image recognition (NMINST).

Main Results:

• The proposed methods significantly outperform the reference method of solely scaling the time constant.
• Achieved substantial accuracy improvements: e.g., 89.5% vs. 53.0% on SHD and 93.6% vs. 38.8% on MSWC when target resolution is double the source.
• Demonstrated that high accuracy on high temporal resolution data is achievable via efficient training on lower temporal resolution data.

Conclusions:

• The novel SSM-based domain adaptation methods effectively address temporal resolution mismatches in SNNs.
• These methods offer a superior alternative to existing techniques, enabling robust SNN deployment across diverse temporal resolutions.
• Efficient training on lower temporal resolutions can yield high accuracy on higher temporal resolution tasks, reducing computational burden.