Approximate, computationally efficient online learning in Bayesian spiking neurons.

Area of Science:

• Computational neuroscience
• Machine learning
• Artificial intelligence

Background:

• Bayesian spiking neurons (BSNs) offer a probabilistic framework for neural inference and learning.
• Current online learning methods like maximum-likelihood expectation-maximization (ML-EM) are computationally intensive, hindering large-scale BSN network studies.

Purpose of the Study:

• To introduce a computationally efficient online learning algorithm for BSNs.
• To compare the efficiency and performance of the new algorithm against ML-EM.
• To assess the suitability of the new algorithm for large-scale BSN network simulations and neuromorphic implementation.

Main Methods:

• Development of the fast learning (FL) algorithm for BSN online learning.
• Comparative analysis of FL and ML-EM in terms of computational speed and convergence.
• Evaluation of FL's robustness to parameter initialization and its estimation accuracy.

Main Results:

• FL demonstrates significantly faster run times than ML-EM, especially with increasing input numbers.
• Despite slower convergence rates, FL's lower computational cost leads to faster overall simulation times to convergence compared to ML-EM.
• FL shows robustness to parameter initialization and good average estimation accuracy within physiologically relevant ranges.

Conclusions:

• The FL algorithm provides a computationally efficient and viable alternative to ML-EM for BSN online learning.
• FL facilitates more detailed exploration of BSN networks and their biological relevance.
• FL's simplicity makes it suitable for implementation in energy-efficient neuromorphic hardware.