Deep learning

Deep learning uses neural networks with many layers that learn their own features from data, instead of relying on features an analyst designs by hand. It is the family of methods behind modern language models, and it is just as much a tool for numbers, time series, sensor signals, and networks of related records.

Where it earns its place

Long or irregular sequences, such as transaction histories or equipment telemetry, where the order of events carries the signal. High-volume sensor data, where hand-built features miss the patterns that matter. Relationship data, such as payment networks, where a graph model can see connections a row-by-row model cannot. Very large datasets, where the extra capacity is actually used.

Where it does not

On a few thousand rows of well-understood tabular data, a scorecard or a gradient-boosted model is often as accurate and far easier to explain, validate, and monitor. A deep model needs more data, more compute, and more work to explain an individual result. Whether its extra lift survives representative out-of-time and segment testing, and earns that burden, is a question for the evidence, not a default.

When there is little data

A network pretrained on large public time-series or tabular data can be fine-tuned or calibrated to an institution's own records, which can make deep learning viable before years of history exist. The result is still evaluated as a new candidate against simpler models.

The machine-learning models page chooses between these methods from the intended use and the evidence; language and text models belong to Small language models.