Feature Extraction

Feature extraction turns raw data into a smaller set of numbers that algorithms can use for decisions or predictions.

The input might be an image, a sound recording, or a spreadsheet of sensor readings. The output is features: measurements that keep the informative parts. In image recognition a camera captures millions of pixels, but a classifier often only needs edges, textures, and shapes to tell a cat from a dog. A smaller descriptor vector is faster, uses less memory, and can raise accuracy.

Speech processing turns raw audio into frequency bands. Finance turns transaction streams into volatility, volume, and trend indicators.

Good features make models faster and more accurate. Bad features make them slow and unreliable. Deep learning can learn features from data, which reduces hand work. In many domains, carefully designed features still beat fully automatic extraction.

Raw pixels, waveforms, or tick streams are high dimensional and redundant. Features are the summary: edges, textures, shapes; frequency bands; volatility, volume, trend. A classifier on those numbers is smaller and often more accurate than one on raw inputs. Pipelines used to hand-design this step.

Deep nets can learn it, but sensor fusion, finance, and some medical tasks still use crafted features when data is limited or when engineers must know what the model looks at. Bad features hide the signal. Good ones make the later model almost easy. Early layers of a trained net become feature extractors. People freeze them and train a small head on a new task.