Representation learning discovers useful features from data instead of hand-designing them

Simple machine-learning algorithms depend heavily on how their input is represented, and choosing good features by hand is slow and hard. Representation learning is the approach of using machine learning to discover not just the mapping from representation to output, but the representation itself. Learned representations often result in much better performance than can be obtained with hand-designed ones, and let systems adapt rapidly to new tasks with little human effort. The autoencoder — an encoder that converts input into a representation plus a decoder that converts it back — is the quintessential example. Deep learning is a form of representation learning where representations are built in layers.

Examples

Instead of an engineer hand-crafting edge/texture features for image recognition, a representation-learning algorithm discovers such features automatically. An autoencoder learns a compressed representation by being trained to reconstruct its own input.

Assessment

Contrast hand-designed features with learned representations, and explain how an autoencoder’s encoder/decoder structure produces a learned representation.