Bayesian continual learning via spiking neural networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Biological intelligence excels in energy efficiency, adaptation, and uncertainty quantification.
• Neuromorphic engineering primarily focuses on energy-efficient, brain-inspired computing.
• Existing systems often lack robust adaptation to new tasks and reliable uncertainty estimation.

Purpose of the Study:

• To develop neuromorphic systems capable of continual learning and accurate uncertainty quantification.
• To design online learning rules for spiking neural networks (SNNs) within a Bayesian framework.
• To enable synaptic weights to represent epistemic uncertainty from prior knowledge and observed data.

Main Methods:

• Derivation of online learning rules for SNNs using a Bayesian continual learning framework.
• Representation of synaptic weights by parameters quantifying epistemic uncertainty.
• Streaming updates of distribution parameters as new data become available.
• Instantiation for both real-valued and binary synaptic weights.

Main Results:

• Demonstrated the capacity for adaptation to changing learning tasks.
• Showcased the ability to produce well-calibrated uncertainty quantification estimates.
• Validated the approach using Intel's Lava platform, comparing Bayesian and frequentist methods.

Conclusions:

• Bayesian learning offers superior adaptation and uncertainty quantification compared to frequentist approaches in SNNs.
• The proposed online learning rules facilitate the development of more intelligent and reliable neuromorphic systems.
• This work advances the design of adaptive, uncertainty-aware neuromorphic engineering.