What Deep Learning Is

Representation learning

Classic machine learning usually relies on features designed by people (word counts, ratios, edges). Deep learning uses neural networks with many layers that learn useful representations directly from raw data: early layers detect simple patterns (edges in an image, character pieces in text), later layers combine them into complex ones (shapes, objects, meanings). This is why deep learning dominates images, audio, text and other unstructured data. For small tabular datasets, tree ensembles are often as good or better and far cheaper, so deep learning is a tool, not a default.

Neurons, layers, depth

A neural network stacks simple units into layers that learn their own features.

Learning to read

A child first learns strokes, then letters, then words, then sentences; each level builds on the one below. Deep networks learn layered features in a similar spirit.

Check whether you need it

For a few thousand rows of tabular data, try gradient boosting first; reach for deep learning when the data is images, audio, text or very large.

Quick check: What is a key advantage of deep learning over classic ML?

• It is always cheaper
• It never needs data
• It learns useful features directly from raw data
• It works best on tiny tables