⌘K
ENES
GitHub
conceptLearning Foundations~1 min readUpdated 2026-06-07#foundations#distribution-shift#drift#robustness

The data distribution & distribution shift

Every model is trained on a sample from some distribution and silently assumes production data comes from the same one. When the real world drifts away from that assumption, accuracy drops — often without any error, just quietly worse outputs.

A vocabulary for what changed

  • Covariate shift — the inputs change, the input→output relationship doesn't. (New user demographics; a camera with different lighting.)
  • Label shift — the mix of outcomes changes. (Fraud base rate jumps during a holiday.)
  • Concept drift — the input→output relationship itself changes. (What counts as "spam" evolves as spammers adapt.) The hardest kind.

Why it's the default, not the exception

  • The world is non-stationary: behavior, language, prices, and adversaries move.
  • Your training set is a snapshot; deployment is a stream.
  • Feedback loops: the model's own actions change the distribution it later sees (a recommender reshapes what users click).

Detecting and responding

  • Monitor inputs, not just outputs — track feature distributions and flag when production data drifts from training (PSI, KS tests, embedding-distance checks).
  • Monitor a proxy for quality when labels are delayed (confidence, human override rate, downstream metrics).
  • Respond with retraining, recency-weighting, or — for LLM systems — refreshing the retrieved context so facts stay current without touching weights.

For LLMs specifically

A model's knowledge cutoff is a built-in distribution shift the day it ships: the world keeps moving, the weights don't. This is the core argument for retrieval over fine-tuning when facts change.

Connects to: realistic splits · monitoring & drift · RAG for fresh facts

Related notes

Learning FoundationsEvaluation metrics & what they hideAccuracy lies on imbalanced data. Precision, recall, F1, ROC-AUC — what each measures, and how to pick the metric that matches the cost of being wrong.Learning FoundationsFeatures, representations & the curse of dimensionalityModels don't see the world — they see features. How representation quality decides everything, and why high dimensions are weird.Learning FoundationsGeneralization, overfitting & the bias–variance tradeoffA model is only useful if it works on data it never saw. Overfitting, underfitting, and the tradeoff that governs both.Learning FoundationsHow learning works: loss, objective, and ERMEvery model that "learns" is minimizing a loss function over data. Understand that loop and most of ML stops being magic.