Lesson 9 of 38 · Written learning
How machine learning learns from examples
Distinguish learning setups and design data splits that make evaluation more honest.
In supervised learning, examples come with target labels. In unsupervised learning, the system looks for structure without those labels. Self-supervised learning creates training targets from the data itself; predicting the next token in a text sequence is one example. Each setup provides a different learning signal, not a guarantee that the result is useful.
Separate fitting from checking
Training data updates the model. Validation data helps compare choices such as model size or learning rate. A test set is held back for a final, less biased estimate. If near-duplicate documents or future information leak across these splits, performance can look impressive without representing real use.
- Features are inputs; labels are targets when labels exist.
- Underfitting means the model has not captured useful structure; overfitting means it fits training details that do not generalize.
- Regularization and careful data collection can reduce overfitting, but neither substitutes for a good test design.
- Data leakage can occur through duplicated records, preprocessing, timestamps, or benchmark answers embedded in training material.
Return to the Machine learning and deep learning stage · Suggested timing is flexible; move at your own pace.