Astrocyte-gated multi-timescale plasticity for online continual learning in deep spiking neural networks

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) offer energy-efficient, event-driven computation ideal for real-time sensory data.
• Training deep SNNs online and continually faces challenges like memory bottlenecks with Backpropagation-through-Time (BPTT) and the stability-plasticity dilemma in local learning rules.

Purpose of the Study:

• To develop a scalable, online learning framework for SNNs that addresses the limitations of existing training methods.
• To introduce a biologically inspired mechanism, Astrocyte-Gated Multi-Timescale Plasticity (AGMP), for robust continual learning in SNNs.

Main Methods:

• AGMP augments eligibility traces with a broadcast teaching signal and an astrocyte-mediated gating mechanism.
• A slow astrocytic variable dynamically modulates plasticity based on neuronal activity, suppressing updates during stable periods and enabling adaptation during distribution shifts.
• The framework was evaluated on neuromorphic benchmarks (N-Caltech101, DVS128 Gesture, SHD) and Class-Incremental Continual Learning tasks (Split CIFAR-100).

Main Results:

• AGMP achieves accuracy comparable to offline BPTT while maintaining constant O(1) temporal memory complexity.
• In Class-Incremental Continual Learning, AGMP significantly reduces catastrophic forgetting without needing replay buffers.
• AGMP outperforms existing state-of-the-art online learning rules in continual learning scenarios.

Conclusions:

• AGMP presents a biologically grounded and hardware-friendly approach for lifelong learning in autonomous agents.
• The proposed method offers a viable solution for robust and efficient online and continual training of SNNs.