Hierarchical Bayesian Inference and Learning in Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Human brain processing aligns with Bayesian principles for environmental complexity.
• Hierarchical Bayesian inference is a leading framework for cortical processing models.
• Neural implementation of hierarchical Bayesian inference in spiking neural networks remains unclear.

Purpose of the Study:

• To propose a biologically plausible spiking neural network architecture for hierarchical Bayesian inference.
• To elucidate the computational mechanisms underlying Bayesian inference in neural networks.
• To demonstrate the network's utility in a practical machine learning task.

Main Methods:

• Developed a hierarchical network of winner-take-all circuits.
• Utilized a spike-based variational Expectation-Maximization (EM) algorithm for inference and learning.
• Mapped neural firing activity to variational E-step and spike-timing-dependent plasticity to M-step.

Main Results:

• The proposed network successfully implements hierarchical Bayesian inference and learning.
• Neural firing patterns correspond to variational inference steps.
• Spike-timing-dependent plasticity models the learning (M-step).
• Demonstrated effective unsupervised classification on the MNIST dataset.

Conclusions:

• The spiking neural network provides a biologically plausible mechanism for hierarchical Bayesian inference.
• This model bridges computational theory and neural implementation.
• The approach shows promise for unsupervised learning in complex environments.