The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / U
3 terms, explained without the techno-murk.
When a model is too simple or insufficiently trained to learn the important pattern.
It is the opposite side of overfitting. Imagine trying to describe a winding coastline with one straight line: the rule is so simple that it performs poorly even on the examples it was given, as well as on new ones.
A basic straight-line model fails to capture a strongly curved relationship in the data.
Information that is not neatly arranged in fixed rows and columns.
Emails, documents, photographs, recordings and videos are commonly called unstructured data. They still have patterns, but need more processing than a tidy database table.
A folder of customer emails is unstructured data.
Learning patterns from data that has not been given human-made answer labels.
The system looks for structure, similarity or useful representations on its own. Much language-model pretraining is often described as self-supervised, a related approach where the data supplies its own learning target.
A system groups news articles by topic without being told the topic names in advance.