Spiking Neural Networks with Adaptive Membrane Time Constant for Event-Based Tracking

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) are brain-inspired models processing data in an event-driven manner.
• SNNs possess inherent temporal memory capabilities suitable for event-based data streams.
• Current SNN applications are mainly focused on classification, with limited exploration in tracking tasks.

Purpose of the Study:

• To investigate the effectiveness of SNNs for event-based tracking tasks.
• To propose novel components for enhancing SNN performance in temporal data processing.
• To demonstrate the adaptability and generalization of the proposed SNN approach.

Main Methods:

• Introduction of a brain-inspired adaptive Leaky Integrate-and-Fire (BA-LIF) neuron.
• The BA-LIF neuron adaptively adjusts its membrane time constant based on input.
• Incorporation of a Temporal Feature Aggregator (TFA) to assign attention weights across the temporal dimension.

Main Results:

• The proposed BA-LIF neuron effectively filters noise and preserves valuable information.
• SNNs utilizing BA-LIF neurons achieve high performance without extensive parameter tuning.
• The TFA further enhances the network's adaptive capabilities for temporal feature extraction.
• Experiments on event-based tracking datasets confirm the method's effectiveness.

Conclusions:

• The developed SNN approach, featuring BA-LIF neurons and TFA, shows significant promise for event-based tracking.
• The adaptive nature of the proposed neurons reduces the need for manual parameter optimization.
• The method demonstrates strong generalization capabilities, applicable to both tracking and classification tasks.