To date, most successful applications of deep learning in signals and decoding are based on supervised learning. However, supervised learning is contingent on the availability of labeled data. The need for labeled data is a serious limitation to applications at scale and complicates the maintenance of real-life supervised learning systems. The typical situation is that unlabeled data is abundant, and this has given rise to paradigms such as self-supervised learning (SSL). SSL is based on combining large amounts of unlabeled data with limited labeled data and relies on supervised learning with a supervisory signal derived from the data itself. While we have seen many successful applications of SSL, theoretical understanding of the mechanisms that make SSL work is only starting to emerge.

The workshop is organized by the Pioneer Centre for Artificial Intelligence’s collaboratory, Signals and Decoding.