Models do not see raw reality. They see features, the measurable properties extracted from data and fed into a learning algorithm. A house becomes square footage, number of bedrooms, age, and location. A document becomes word counts or embeddings.

Feature engineering once dominated applied machine learning. Practitioners hand-crafted ratios, log transforms, and interaction terms to help models find patterns. Deep learning shifted the work, letting networks learn features directly from raw input.

Common feature types

• Numerical values like price or temperature
• Categorical labels encoded as numbers
• Text converted to counts or vectors
• Image pixels or learned embeddings
• Derived ratios and interactions

Good features carry signal. Redundant or noisy ones slow training and hurt generalization. Feature selection and extraction remain important, especially on small datasets where deep learning has little advantage.