LeCun weighs in on bioplausible learning debate: "everything works on MNIST"
aran_nayebi · x · 2026-09-16
Sakana AI's bioplausible learning work drew in Yann LeCun, who noted it is essentially a version of "target prop": each layer's input acts as a free latent variable serving as a target for the previous layer, derivable from an augmented Lagrangian formulation of backprop where layer-consistency constraints become penalty terms.
Researcher Aran Nayebi pushed back: these results are only on MNIST, where even trivial methods like Feedback Alignment work — "everything works on MNIST." The real test for a bioplausible algorithm is whether it learns at scale without weight transport; he points to work from six years ago showing bioplausible learning scaling with backprop on ImageNet, and links the relevant paper.
Related event: Sakana AI Unveils PC-ALM: Training 1000-Layer Nets Without Backprop(5 posts)→
More from Research
- Researchers ask: why is there no AI-first scientific journal with AI-run peer review? — anshulkundaje · 2026-09-21
- Waypoint Bio argues AI in biology has a data problem before a model problem — anshulkundaje · 2026-09-21
- New paper estimates LLMs raised US natural unemployment rate by 0.1-0.2 points — mattbeane · 2026-09-21
- JevBench open-sources a benchmark for typed decision models scoring intelligence, calibration and speed — sull · 2026-09-21
- gLM2 designs a 2,500-AA polyketide synthase making a nylon precursor, 10x titer boost — anshulkundaje · 2026-09-21
- Skip the LLM for routing? Jev classifier ties Llama 70B in harness tests — 2BucChuck · 2026-09-21