-Illustration of the key characteristics of the neural oscillation mechanism

 

 

Mr. Yinsong Yan, a second-year PhD student from the MIND Lab, has had his research paper "Efficient and Robust Temporal Processing with Neural Oscillations Modulated Spiking Neural Networks" accepted for publication in Nature Communications. Nature Communications is a highly regarded multidisciplinary scientific journal published by the Nature Portfolio since 2010. This selective publication covers research across the natural sciences, including physics, chemistry, earth sciences, medicine, and biology, and is recognized for its rigorous peer-review process and impact on the scientific community. The journal consistently ranks among top-tier publications in scientific research, representing a notable achievement for the research team.

 

This study, co-authored with Dr. Yujie Wu, Prof. Kay Chen Tan, and Dr. Jibin Wu from PolyU, alongside international collaborators Dr. Qu Yang and Prof. Haizhou Li from National University of Singapore, and Mr. Hanwen Liu and Dr. Malu Zhang from University of Electronic Science and Technology of China, introduces a novel computational framework inspired by neural oscillations in the human brain. The research proposes incorporating rhythmic neural modulation into spiking neural networks (SNNs) to enhance their temporal processing capabilities. By introducing heterogeneous oscillatory signals that modulate spiking neurons to activate periodically at distinct frequencies, the authors have developed "Rhythm-SNNs." This approach significantly reduces neural firing rates while improving the capability and robustness of SNNs in temporal processing. Extensive experimental results demonstrate that Rhythm-SNNs achieve significant improvements in temporal processing capability, energy efficiency, and robustness against perturbations. Furthermore, this research provides theoretical analysis demonstrating computational advantages in three key areas: effective gradient backpropagation pathways, enhanced memory capacity, and improved robustness against perturbations.

 

This work demonstrates promising applications in neuromorphic computing, particularly in temporal signal processing. Notably, in the Intel Neuromorphic Deep Noise Suppression Challenge, the Rhythm-SNN achieved award-winning denoising performance while delivering over two orders of magnitude improvement in energy efficiency compared to traditional deep learning solutions. The research opens possibilities for developing next-generation neuromorphic hearing devices, including hearing aids and headsets that can operate efficiently in complex environments. Looking ahead, the MIND Lab team will continue exploring biological working mechanisms underlying neural systems, aiming to develop more efficient and robust neuromorphic signal processing systems capable of deployment on resource-constrained devices in real-world scenarios.