Gary Marcus: universal approximation doesn't mean networks can learn
GaryMarcus · x · 2026-11-22
Gary Marcus endorses Chomba Bupe's point that the universal approximation theorem only says a neural network can approximate any function in principle — not that it can learn any function from input-output pairs.
Marcus claims he made exactly this argument 25 years ago in his first book and that its importance is still barely understood, continuing his critique of deep learning's theoretical foundations.
More from AGI Musings
- Gary Marcus hails 'centaur math' milestone as Lean certifies human+AI proof — GaryMarcus · 2026-11-22
- Yoav Goldberg: AI safety community must publicly distance itself from its most extreme voices — yoavgo · 2026-11-22
- Gary Marcus: LLMs work better because they borrowed tree search from symbolic AI — GaryMarcus · 2026-11-22
- A mathematician despairs when the machine solves it first; an engineer moves on — yunta_tsai · 2026-11-22
- LeCun: Training LLMs on human knowledge yields memory, not intelligence — haider1 · 2026-11-22
- David Deutsch: Explanation, not proof, is the point of pure mathematics — TacoCohen · 2026-11-22