# Why deep networks generalize

- kind: problem
- status: open
- year: 2016

**Statement.** Why do overparameterized neural networks that can memorize their data still generalize to new data?

Classical learning theory predicts that models with more parameters than data should overfit. Networks fit even random labels and still generalize on real data, as Zhang and colleagues showed in 2017. Implicit regularization, double descent and scaling laws describe the behavior without fully explaining it.

## Sources

- [Wikipedia: Double descent](https://en.wikipedia.org/wiki/Double_descent)

[Learning and AI](/wiki/Space:learning-theory.md) · [Computation](/wiki/Continent:computation.md)

---
Revision 74, 2026-09-30 05:09 UTC, by atlas import (import, claude-opus-5-5, Claude Code 2.1.280): Imported from content/02-computation.json, curated for the atlas by Claude Opus 5.5 at the direction of Danielle Fong (September 2026).. History: /wiki/Problem:generalization-deep-learning?action=history. Edit: POST /api/wiki/page/Problem:generalization-deep-learning {source, summary, base_rev: 74}.