Double Descent
Updated 2026-08-03
INTRODUCTION
English translation pending.
CORE DEFINITION
Double descent is a phenomenon in modern statistical learning in which test error fails to follow the textbook U-shaped bias-variance curve. As model complexity grows, error first falls, then rises toward a peak near the interpolation threshold where the model just fits the training set, and then falls a second time as parameters keep multiplying. The result overturns the classical Occam's razor instinct in artificial intelligence, because in the over-parameterized regime, with abundant data, extra capacity proves harmless and even becomes the route to better generalization.
SCAFFOLDING EFFECT
Reduce cognitive load
- Curve check: Plot train and test error together and look for the second drop instead of stopping at the first rise. - Capacity patience: Do not halt tuning just because classical theory says the model is now overfit. - Regime locator: Use a validation set to find which side of the interpolation threshold your setting sits on.
Anchor fast decisions
A model with more parameters than samples can interpolate the training set exactly, which classical theory predicts should wreck generalization. Yet past that threshold, extra parameters let the model fit the data with smoother, lower-norm solutions that average out noise. So error, after its intermediate peak, declines again as capacity grows, and over-parameterization becomes an advantage rather than a liability.
MINIMUM ACTION
In progress 0/1Practice this model in one real situation:
account_treeGenealogyexpand_more
menu_bookReferencesexpand_more
Source support: Explicit
- en.wikipedia.orghttps://en.wikipedia.org/wiki/Double_descentverified
PRIVATE NOTES · Only visible to you
SAVED Q&A
ENTRY Q&A · Private saving available
Ask with a clear boundary
thinkingmodels answers from published entry context only.
Your question is sent to thinkingmodels. The answer uses public entry context only.
RELATED MODELS