Backpropagation through space, time and the brain

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Neuronal networks face challenges in credit assignment due to spatio-temporal locality constraints.
• Existing backpropagation algorithms often violate these locality principles.

Purpose of the Study:

• Introduce Generalized Latent Equilibrium (GLE) for fully local spatio-temporal credit assignment.
• Develop a framework for efficient learning in physical, dynamical neuronal networks.

Main Methods:

• Derived neuronal dynamics from an energy function based on local mismatches.
• Utilized stationarity and gradient descent for parameter dynamics.
• Exploited dendritic morphology for information processing and prospective coding.

Main Results:

• Developed an online approximation of backpropagation through space and time.
• Demonstrated computation of spatio-temporal convolutions in the forward direction.
• Showcased approximation of adjoint variables in the backward stream.

Conclusions:

• GLE provides a biologically plausible mechanism for credit assignment in neuronal networks.
• This framework supports continuous learning with local synaptic plasticity.
• Prospective coding enhances computational capabilities within single neurons.