PNAS Nexus · 23 September 2026 · From issue 41
Deep learning without weight symmetry
A new learning rule trains networks well without needing matched two-way wiring between neurons.
Li Ji-An, Marcus K Benna.
Computer programs called neural networks usually learn using a method that brains probably do not use. That method needs matching wiring going both ways between brain cells. But brain studies show many connections only go one way. A new learning rule called Product Feedback Alignment lets networks learn well without that matching wiring. It matched the performance of the older method on deep image recognition networks.
Integrity screen: passed (3 checks) Checked 6 October 2026. Read the source
The plain-language reading of this paper is in the full issue, for subscribers.
Read the full issue