Improving Machine learning

The ``self-supervised'' work makes two important contributions: It offers a framework for thinking about the interaction between sensory modalities and the role of feedback in learning, and it has the potential to vastly improve machine learning algorithms, particularly opening up the field to the use of unsupervised data. The major drawback with traditional supervised machine learning algorithms is that they require a lot of expensive hand-labeling of training patterns in order to generalize well. Providing alternative ``self-supervised" algorithms which supervise themselves through richer use of the input, will allow collection of much more training data (for example through the internet where information is abundant but unlabeled).