STDP-Compatible Approximation of Backpropagation in an Energy-Based Model.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models often use backpropagation for efficient credit assignment.
• Energy-based models with latent variables present challenges for gradient propagation.
• Understanding biological learning mechanisms, like credit assignment in brains, is crucial for AI advancement.

Purpose of the Study:

• To investigate Langevin Markov chain Monte Carlo (MCMC) inference in energy-based models.
• To explore the connection between MCMC inference and backpropagation in deep neural networks.
• To propose a theoretical framework for brain-like credit assignment in artificial systems.

Main Methods:

• Utilized Langevin MCMC inference within an energy-based model framework.
• Analyzed the early stages of inference starting from a stationary point.
• Developed theoretical arguments and simulations to validate findings.

Main Results:

• Early MCMC inference steps propagate error gradients into internal layers, mirroring backpropagation.
• Backpropagated errors are linked to output units driven away from a stationary point.
• Derived a rate-based weight update rule consistent with spike-timing-dependent plasticity.

Conclusions:

• Langevin MCMC inference offers a mechanism for approximate credit assignment in deep hierarchies.
• Neural computation can simultaneously perform approximate inference and error backpropagation.
• This approach provides insights into how brains might achieve efficient credit assignment, similar to backpropagation.