Researchers have developed a learning mechanism that uses the natural variability of neural activity — often dismissed as random “noise” — to understand how synapses buried deep inside brain networks — or brain-inspired devices — can adapt to improve the learning capabilities of the networks as a whole.
To do so, they needed to solve the so-called weight transport problem, a missing piece in many models of deep learning in spiking neuronal networks.
Modern AI universally relies on the error backpropagation algorithm, which adjusts connections between artificial neurons during learning. But these algorithms make an assumption that is difficult to reconcile with how real brains work: information travelling forward (sensory representations) and backward (response errors) in the network must essentially take the same path. Because connections between neurons are unidirectional, these two signals need to travel through separate, but otherwise identical pathways. In a computer, copying this information between parts of a network is trivial. But brains rely on local interactions between neurons and synapses — there’s no known way for one synapse to “read off” the exact strength of another, distant one.
Neuromorphic chips, which mimic neurons and synapses directly in hardware for more energy-efficient AI, face the same obstacle. Solving this weight transport problem represents an important step toward AI that combines the power of error backpropagation with the efficiency and biological realism of brain-inspired computing.
Aligning connections without copying them
In a study published in Nature Communications, researchers at the University of Bern, with colleagues at Heidelberg University and the Okinawa Institute of Science and Technology, introduce Spike-based Alignment Learning (SAL) — a learning rule that solves the weight transport problem without ever copying connection information across the network.
Instead, SAL puts the messy, variable timing of neurons’ electrical pulses, or spikes, to work. As two connected neurons fire, tiny, naturally occurring differences in exactly when their spikes occur carry a subtle signal about how mismatched their forward and backward connections are. By tracking these timing differences, each connection can gradually correct itself using only information available locally, at that single synapse. The randomness dismissed as unwanted noise turns out to carry exactly the information needed to keep a network’s forward and backward pathways in sync.
The team tested SAL across several brain-inspired network types — spiking networks for probabilistic reasoning, biologically plausible variants of error-driven learning, and, as a benchmark, a deep neural network on an image classification task. In each case, SAL matched the performance of existing approaches while avoiding their unrealistic requirement that connection information be shared across the network. It also proved robust to a challenge shared by biological brains and analogue computer chips alike: no two neurons or synapses are ever perfectly identical. Because SAL continuously adapts to these differences, it could help future neuromorphic systems stay reliable despite the manufacturing imperfections and device-to-device variability unavoidable in physical hardware.
Two of the study’s authors are supported by the EBRAINS 2.0 project, which further develops EBRAINS data tools and services — including the kind of brain-inspired hardware and spiking neural network models SAL was designed for.
Original Publication
Gierlich, T., Baumbach, A., Kungl, A.F., Petrovici MA. Spike-based alignment learning solves the weight transport problem. Nat Commun (2026). https://doi.org/10.1038/s41467-026-74460-8
Author: Helen Mendes
Contact: press@ebrains.eu
Create an account
EBRAINS is open and free. Sign up now for complete access to our tools and services.