# Local Learning Rules Challenge Backpropagation at Depth

Training methods that replace global backpropagation with local, per-layer credit assignment keep closing the accuracy gap at increasing depth, so expect recurring results scaling local-rule training to more layers and harder datasets, and memory-locality arguments — not accuracy alone — driving interest from hardware vendors whose accelerators are bottlenecked by storing activations for the backward pass.

- Conviction: 34 / 100 (forming)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-09-15T00:00:00.000Z
- Last updated: 2026-09-15T06:50:06.768Z
- Canonical: https://polylog.news/ai/trends/backprop-free-deep-network-training
- Publisher: Polylog
- Affected regions: Global

## Recent evidence

- [confirms] Sakana AI trains 1,000-layer networks without backpropagation (2026-09-15): Sakana AI trained 1,000-layer networks without backpropagation by accumulating per-layer constraint errors into a local Lagrange multiplier, landing within about two points of backpropagation accuracy on MNIST at that depth. Depth of 1,000 layers is far beyond where prior local-learning methods degraded, though the result is still on a toy dataset.