Research‎ > ‎

Representation learning

Representation is a fundamental process for machine learning. Its goal is to extract useful features from training data which are later fed to a learning algorithm. In computer vision, representation corresponds to calculate values from input images. In audio signal processing, representation corresponds to compute numerical values that characterizes the complex nature of audio signals. These values (features) represent particular characteristics of the original data and are calculated from the raw values. The functions used to compute such features are called feature detectors.
Traditional approaches are based on standard or hand-crafted feature detectors which are manually selected to fit the problem at hand using expert knowledge in the domain. A main drawback in hand-crafted features is the high cost of such expert intervention. Experts usually have to design a different set of features for each problem. Representation learning tackle this problem from a different perspective. Instead of designing custom feature detectors, representation learning learns them from data.