Temporal dendritic heterogeneity incorporated with spiking neural networks for learning multi-timescale dynamics

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) show promise for temporal information processing due to their dynamic nature.
• Understanding the learning mechanisms and exploiting the dynamic properties of SNNs for complex temporal tasks remains a challenge.

Purpose of the Study:

• To propose a novel multi-compartment spiking neural model capable of capturing multi-timescale temporal dynamics.
• To investigate the underlying mechanisms of learning and temporal feature integration in SNNs.
• To demonstrate the practical benefits of the proposed model on diverse temporal computing benchmarks.

Main Methods:

• Development of a multi-compartment spiking neural model incorporating temporal dendritic heterogeneity.
• Automatic learning of heterogeneous timing factors across different dendritic branches to enable multi-timescale dynamics.
• Experimental validation using the temporal spiking XOR problem and benchmarks for speech, visual, electroencephalogram (EEG) signal, and robot place recognition.

Main Results:

• The working mechanism of temporal feature integration at different levels was elucidated using the temporal spiking XOR problem.
• The proposed model significantly outperformed ordinary spiking neural networks on multiple temporal computing tasks.
• Achieved state-of-the-art accuracy, improved model compactness, robustness, generalization, and high execution efficiency on neuromorphic hardware.

Conclusions:

• The proposed multi-compartment SNN model effectively captures multi-timescale dynamics by learning heterogeneous timing factors.
• This approach enhances performance in complex temporal computing tasks, including speech, visual, and EEG recognition.
• The findings represent a significant advancement for neuromorphic computing, bringing it closer to real-world applications by leveraging biological insights.