Generalization
In machine learning, generalisation refers to how well a trained model performs on unseen data. A model that generalises effectively captures underlying patterns rather than memorising examples.
Illustration
Imagine a spam detection system: if it only memorises the specific emails from the training set, it will fail on new phishing messages. But if it has learned the statistical cues of spam (suspicious links, unusual wording), it will perform well on previously unseen emails.
Key aspects
- Good generalisation = low error on both training and test sets.
- Poor generalisation arises from overfitting (too complex, memorising noise) or underfitting (too simple, missing patterns).
- Strategies to improve it include cross-validation, regularisation, and larger, more diverse datasets.
Why it matters
Generalisation is central to deploying AI in real-world applications—from medical diagnostics to financial forecasting—because systems will always face data shifts compared to training.
Generalisation is often described as the true test of intelligence in machines: the ability to perform well not just on what has been seen, but on what is new. In practical terms, it means that a model has learned patterns that extend beyond the quirks of the training data.
In modern AI, generalisation has become harder because models are trained on massive datasets that may not fully represent future conditions. For instance, a medical imaging model trained on one hospital’s scans may struggle when deployed in another with different equipment. This is why researchers now study domain adaptation and robust generalisation under distribution shifts.
There is also an ongoing debate: large foundation models appear to generalise impressively across tasks, but sometimes this hides spurious correlations—patterns that work in training but fail in the real world. Ensuring genuine generalisation requires not only data and algorithms, but also careful evaluation in realistic scenarios.
Further reading:
- Andrew Ng, Machine Learning Yearning (2018).
- Bishop, Pattern Recognition and Machine Learning (2006).
