Efficient and robust temporal processing with neural oscillations modulated spiking neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Neuromorphic Engineering

Background:

• The brain's temporal processing relies on complex dynamics, a capability not fully replicated by current spiking neural networks (SNNs).
• Existing SNNs struggle with temporal tasks and are susceptible to noise, limiting their practical applications.
• Neural oscillations are a key mechanism in biological brains for efficient information processing.

Purpose of the Study:

• To enhance the temporal processing capabilities and noise robustness of SNNs.
• To develop a novel SNN architecture inspired by biological neural oscillations.
• To reduce the energy consumption of SNNs for efficient neuromorphic computing.

Main Methods:

• Introduced Rhythm-SNN, a novel SNN architecture.
• Employed heterogeneous oscillatory signals to modulate spiking neuron activation frequencies.
• Conducted extensive experiments and theoretical analyses on various temporal processing tasks.

Main Results:

• Rhythm-SNN significantly reduced neuronal firing rates and enhanced temporal processing capabilities.
• The proposed model demonstrated superior robustness against noise perturbations.
• Achieved state-of-the-art performance across multiple tasks with markedly reduced energy costs.
• Outperformed deep learning solutions in a neuromorphic noise suppression challenge, achieving significant energy reduction.

Conclusions:

• Rhythm-SNN effectively leverages neural oscillation principles to overcome limitations in traditional SNNs.
• The approach offers a promising direction for developing more efficient and robust neuromorphic systems.
• Rhythm-SNN presents a significant advancement in energy-efficient AI for temporal processing and noise suppression.