Representation learning is an unsupervised technique in which a model learns useful encoded features from raw, unlabeled data. It treats the data as a mixture of useful signal and distracting noise, learning a representation that emphasizes the former. section 1.2.2
Representation learning is closely related to several unsupervised-learning approaches:
- Dimensionality reduction preserves important features while discarding less useful information, often making data easier to visualize in two or three dimensions. section 1.2.2
- Compression encodes data in a lower-dimensional form while retaining the possibility of reconstructing the original data. section 1.2.2
- Clustering and generative modeling use learned structure to group data or model the distribution from which it was generated. section 1.2.2




