Feature learning (representation learning) refers to techniques by which a system automatically discovers the representations needed for feature detection or classification from raw data. In classical ML, a lot of effort went into manual feature engineering. Feature learning, especially with deep learning, means the model itself learns good features as part of training. For instance, a CNN automatically learns edge detectors, texture detectors, object part detectors in successive layers without them being hand-specified. Unsupervised feature learning methods like autoencoders, restricted Boltzmann machines, or self-supervised learning can learn latent features from unlabeled data which can then be used for a task. The success of deep learning in images, audio, and text largely comes from automating feature learning – the network figures out intermediate representations that are optimal for the end task, often yielding better performance than hand-crafted features.