Skip to content

Bias/variance, overfitting and regularisation

  • ML fundamentals
  • Asked often
  • About 1 h 30 min to learn
  • 3 vetted resources
Build a plan that includes this

What this topic covers

A model's test error splits into bias (it is too simple to capture the pattern), variance (it changes too much when the training data changes) and irreducible noise. Overfitting is the high-variance failure where training loss keeps falling while validation loss rises; regularisation (L2 penalties, early stopping, more data, simpler models) trades a little bias for a lot less variance. Interviewers ask candidates to diagnose a described loss curve or learning curve, name which lever to pull and why, and explain what the regularisation strength controls. For LLM roles the same ideas return as fine-tuning questions: why a small fine-tune overfits, how to spot it, and when more data beats a bigger model.

Builds on

You’re ready when you can

What interviewers check, from the resources below.

  • Diagnose overfitting vs underfitting from a described pair of training and validation loss curves
  • Explain what the regularization rate (lambda) in an L2 penalty controls and what happens at both extremes
  • Explain early stopping as a form of regularisation and when you would prefer it
  • List three fixes for overfitting and say which one you would try first and why
  • Define bias and variance in terms of repeated training on different samples, not just “too simple / too complex”
  • Draw the U-shaped test-error curve against model complexity and label where each term dominates
  • Explain what double descent is and why it complicates the classic picture for large models
  • Give one concrete example each of a high-bias and a high-variance model
  • Explain the difference between a validation curve and a learning curve and what question each answers
  • Read a learning curve and say whether more data would help or the model is bias-limited

What to study, in order

Each link opened and checked by hand; the date says when.

  1. Start here

    Datasets, generalization, and overfitting (Machine Learning Crash Course) (opens in a new tab)

    Google for Developers · Google · 45 min · Free · checked 21 Oct 2026

    Why
    Interviewers expect you to read a loss curve on sight and name the fix; this module drills exactly that with short lessons and a quiz.
    What to read
    Do the sub-pages “Generalization”, “Overfitting”, “Model complexity”, “L2 regularization” and “Interpreting loss curves”, then “Test your knowledge”.
    What to skip
    The earlier data sub-pages (“Data characteristics”, “Labels”, “Imbalanced datasets”, “Dividing the original dataset”, “Transforming data”) unless you have never split a dataset before.

  2. Free alternative

    The Bias Variance Tradeoff (opens in a new tab)

    MLU-Explain (Amazon Machine Learning University) · Jared Wilber & Brent Werness · 25 min · Free · checked 21 Oct 2026

    Why
    “Explain bias vs variance” is a near-certain fundamentals question, and this visual walk-through gives you a picture you can draw on a whiteboard.
    What to read
    Scroll from “A Simple Model” through “The Bias Variance Trade-off”; play with the “LOESS Regression” slider long enough to see variance change when you randomise the data.
    What to skip
    “K-Nearest Neighbors” if short on time. Read “What About Double Descent?” only for the one-line caveat that the U-shaped curve is not the whole story for large neural nets.

  3. Go deeper

    Validation curves: plotting scores to evaluate models (scikit-learn user guide 3.5) (opens in a new tab)

    scikit-learn documentation · scikit-learn developers · 20 min · Free · checked 21 Oct 2026

    Why
    The practical follow-up to “it overfits” is “would more data help?”, and learning curves are the tool interviewers expect you to reach for.
    What to read
    Read the unnumbered introduction (the bias/variance/noise decomposition with the polynomial example), then “3.5.1. Validation curve” and “3.5.2. Learning curve”.
    What to skip
    The API details of the Display classes; know that validation_curve and learning_curve exist and what they plot, not their argument lists.

Companies that test ML fundamentals most

Put Bias/variance, overfitting and regularisation in your plan

Tell us your interview date and hours a day; it lands on the right day, in the right order.

Build my plan