Sakana's PC-ALM trains 1000-layer nets without backprop; LeCun calls it target prop reborn
ylecun · x · 2026-09-15
Sakana AI introduces PC-ALM, a local-learning alternative to backpropagation that trains 1000-layer neural nets using only local dynamics.
- Each layer's input acts as a free latent variable serving as a target for the previous layer — a "target prop" approach
- It can be derived from an augmented Lagrangian formulation of backprop, turning layer-to-layer consistency constraints into penalties
- Yann LeCun welcomed the revival of an idea his lab explored with sparse auto-encoders in the late 2000s, but noted target prop optimizes the same criterion as backprop — only the gradient evaluation differs, perhaps more biologically plausible
More from Research
- Dynin-Robotics: omnimodal unified diffusion vision-language-action model for robot control — snu-aidas · 2026-09-15
- E2E driving's inflection point: switching from trajectory regression to trajectory classification — abursuc · 2026-09-15
- End-to-End Driving Has Come a Long Way from Single-Mode L2 Imitation — abursuc · 2026-09-15
- Oxford report: many international AI agreements are verifiable without speculative tech — davidmanheim · 2026-09-15
- End-to-end autonomous driving: from direct prediction to counterfactual deliberation at SSAD 2026 — abursuc · 2026-09-15
- Five exercises find four defects in the author's own LLM-judge eval system — alexpran · 2026-09-15