Representation learning
Representation learning trains a model to construct features from data rather than relying entirely on hand-designed features. The learned representation can make later tasks such as classification or retrieval easier.
Also known as: feature learning
A defect-inspection pipeline could train an image encoder and then feed its vectors to a small classifier. Instead of explicitly programming every texture or edge measurement, training shapes the encoder’s internal features. Whether those features make defects easier to distinguish is something to test, not an automatic consequence of producing vectors.
The encoder can learn through labeled classification, self-supervised learning or reconstruction with an autoencoder. Representation learning describes what is learned, while these terms describe particular objectives or model designs.
Reusing the encoder for another task is transfer learning. Evaluate its features on that task: a representation useful for broad object categories may ignore a tiny surface crack. A compact or visually neat embedding space is not by itself evidence that the information needed for a decision has been retained.
Sources
- Bengio, Courville and Vincent: Representation Learning — Reviews learned features, explanatory factors, autoencoders and deep representations for downstream learning.
Go deeper
- Keras: Convolutional autoencoder for image denoising docs
Follow a reconstruction objective that learns internal features without digit-category labels.