Researchers propose biologically plausible approximation of backpropagation through time
An arXiv paper introduces a formalism that approximates BPTT with local, time-continuous learning rules, building on energy-based principles and correcting a gain error in prior work.
In plain words
Training neural networks over time usually relies on backpropagation through time, a method that does not map cleanly onto how brains learn. The paper derives a mathematically controlled approximation of that method using local rules, meaning neurons and synapses adjust using only nearby information in space and time. The authors say this yields a framework for spatiotemporal learning and a blueprint for physical circuits.
What was announced
The authors, writing on arXiv, note that deep learning algorithms remain largely at odds with current understanding of brain circuitry, with backpropagation through time (BPTT) a prominent example. Their formalism starts from a prospective energy function of neuronal states and derives real-time error dynamics for time-continuous networks, which they present as a straightforward derivation of the adjoint method, the time-continuous equivalent of BPTT. They build on principles of energy conservation and extremal action, unifying and extending earlier approaches to local, phase-free spatiotemporal credit assignment.
Because avoiding unphysical temporal non-localities requires further approximations, the paper addresses a weakness in Generalized Latent Equilibrium (GLE), where enforcing temporal locality induces a gain error that becomes problematic for complex signals. The authors design a learning rule for backward weights that corrects this gain difference, producing fully local equations for neuron and synapse dynamics. They position the result as a rigorous framework for spatiotemporal deep learning in the brain and a blueprint for physical circuits performing such computations.
What it does not show
- The source is the paper's abstract only; no experimental results, benchmarks, or independent evaluation are included, so the claimed corrections and framework are not empirically demonstrated in the supplied evidence.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.