Probabilistic inference of binary Markov random fields in spiking neural networks through mean-field approximation

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Cognitive processes are increasingly understood as probabilistic inference.
• Probabilistic graphical models, such as Markov random fields, are used to model these processes.
• Implementing probabilistic inference in spiking neural networks remains a challenge.

Purpose of the Study:

• To propose a spiking neural network model for inference in arbitrary binary Markov random fields.
• To demonstrate the mathematical equivalence between the proposed network's dynamics and Markov random field inference.
• To unify and extend previous approaches.

Main Methods:

• Design of a spiking recurrent neural network.
• Application of mean-field theory to establish mathematical equivalence.
• Theoretical analysis and experimental validation.

Main Results:

• The proposed spiking neural network can implement inference for any binary Markov random field.
• The mean-field approach unifies existing methods.
• The model achieves results comparable to traditional mean-field inference, demonstrated by image denoising applications.

Conclusions:

• A novel spiking neural network model successfully implements general binary Markov random field inference.
• The mean-field approach provides a unified framework for such implementations.
• The model shows practical utility in tasks like image denoising.