Feature extraction is the process of deriving new features from raw data to reduce dimensionality while preserving relevant information. Instead of using raw inputs (which might be high-dimensional and noisy), one computes transformations that distill the data. For example, in image processing, one might extract HOG (Histogram of Oriented Gradients) features from images to feed a classifier, rather than raw pixels, capturing texture/shape information in a compact form. In NLP, turning sentences into TF-IDF weighted word frequencies or word embeddings is feature extraction. This often results in a fixed-length numeric feature vector that can be used by ML algorithms. Feature extraction can be manual (based on domain knowledge) or automated (using techniques like PCA, deep autoencoders, or pre-trained CNNs for images that output high-level features). It’s especially important when dealing with unstructured data like images, text, or audio.