Multisample Online Learning for Probabilistic Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) mimic biological brains for efficient, event-driven processing.
• Existing SNN learning algorithms often use deterministic models and heuristic approximations.
• Probabilistic SNNs offer principled online, local learning rules, ideal for resource-constrained systems.

Purpose of the Study:

• Investigate the advantage of probabilistic SNNs in generating independent outputs for the same input.
• Demonstrate how multiple output samples enhance decision robustness and uncertainty quantification.
• Explore the use of probabilistic SNNs for more accurate training via improved log-loss and gradient estimation.

Main Methods:

• Introduced a novel online learning rule, Generalized Expectation-Maximization for SNNs (GEM-SNN).
• GEM-SNN follows a three-factor form with global learning signals.
• Evaluated GEM-SNN on structured output memorization and neuromorphic classification tasks.

Main Results:

• Probabilistic SNNs generate multiple, independent outputs, unlike deterministic models.
• Increased sample usage during inference improved decision robustness and uncertainty quantification.
• GEM-SNN training with more samples led to significant gains in log-likelihood, accuracy, and calibration.

Conclusions:

• Probabilistic SNNs provide inherent advantages for robust inference and uncertainty estimation.
• The proposed GEM-SNN learning rule effectively leverages multiple samples for improved performance.
• This approach enhances the capabilities of SNNs for complex machine learning tasks.