Sakana AI Unveils PC-ALM: Training 1000-Layer Nets Without Backprop

On September 15, Sakana AI released the paper "Augmented Lagrangian Predictive Coding" (PC-ALM), a local-learning alternative to backpropagation that trains 1000-layer networks using only local interactions between layers, motivated by the observation that the brain cannot precisely implement backprop. The work quickly drew community discussion, with Yann LeCun and other researchers weighing in on its novelty and generalization.

Confirmed

• Sakana AI officially released PC-ALM (m5 is the original post); the method is built on predictive coding and trains 1000-layer networks with purely local dynamics.
• Core result: on MNIST, a 1000-layer residual network trained via local learning lands within about two percentage points of backpropagation accuracy (relayed by demianai).
• Yann LeCun commented that the method is essentially a variant of target propagation—each layer's input acts as a free latent variable providing targets for the previous layer—and can be derived from backprop with an augmented Lagrangian (relayed by arannayebi).
• Researcher Pierce Minervini responded that the idea is not new, with similar work dating to around 2018 (m3).

Unconfirmed

• Performance on larger datasets and deeper/wider networks; current materials only cover MNIST results.

Why it matters

• Backpropagation underpins deep learning; if local learning can approach its accuracy, PC-ALM could inform brain-like learning mechanisms and reduce global gradient communication costs.
• However, as LeCun quipped that "everything works on MNIST," and given the prior work cited by Minervini, the method's novelty and scalability still require stronger evidence.

2026-09-15 ~ 2026-09-16 · 5 related posts

• [source] Sakana AI's PC-ALM trains 1000-layer nets without backprop; researcher notes 2018 precedent — PMinervini · 2026-09-15
• Sakana's PC-ALM trains 1000-layer nets without backprop; LeCun calls it target prop reborn — ylecun · 2026-09-15
• Sakana's PC-ALM Trains 1,000-Layer Networks Without Backprop, Within ~2 Points of SGD — demian_ai · 2026-09-15
• [source] Sakana AI's PC-ALM trains 1000-layer nets with local learning, no backpropagation — SakanaAILabs · 2026-09-16
• [source] LeCun weighs in on bioplausible learning debate: "everything works on MNIST" — aran_nayebi · 2026-09-16