Overfitting, Underfitting and Generalisation: Why a Perfect Training Score Can Fail

By the end, you can explain why a model that is perfect on its training data can fail on new data, spot underfitting and overfitting from a training error and a test error, and name the usual remedies. We fit a straight line, a degree-3 curve and a degree-15 curve to the same 10 noisy points and compare their errors on 50 held-back test points. Chapters 00:00 Last time: five terms 00:34 Two students, two exam papers 01:41 Generalisation and the test set 02:34 Our model: a polynomial 04:03 Worked example: degrees 1, 3 and 15 05:33 Quick quiz: read the two numbers 06:12 The same experiment in NumPy 06:37 How to fight overfitting 07:53 Two interview questions 08:18 Three things to remember Note: the noisy points and all error numbers come from a toy experiment made for this video (fixed random seed), not from a benchmark. Narration is a synthetic (AI-generated) voice, cloned from the creator's own voice with their consent; script and visuals were prepared with AI tools and reviewed by a human. Sources https://scikit-learn.org/stable/auto_... https://scikit-learn.org/stable/modul... https://www.deeplearningbook.org/cont... https://hastie.su.domains/ElemStatLearn/ https://arxiv.org/abs/1812.11118 https://arxiv.org/abs/1912.02292 https://scikit-learn.org/stable/modul... https://scikit-learn.org/stable/modul... https://scikit-learn.org/stable/modul... https://scikit-learn.org/stable/auto_... https://en.wikipedia.org/wiki/Bias%E2... https://en.wikipedia.org/wiki/Early_s... Previous: P0-06 — Data, Features, Labels, Parameters, Hyperparameters Next: P0-08 — A Short History of AI: Perceptron to Foundation Models