Deep learning with differential equations

Machine learning is developing at an unprecented pace due to a paradigm shift caused by deep neural network models, which have revolutionised the several domains of science. Deep neural networks represents learning as a series of deterministic, complex and discrete transformations. In this Aalto University research project we will propose a groundbreaking new viewpoint on machine learning by developing a novel deep learning paradigm of probabilistic continuous-time deep learning, where interpretable, simple distributions of smooth transformations, or time differentials, encode the learning process as a continuous flow. The novel paradigm draws from solid foundations of physics, statistics and dynamical systems literature. The project will be performed in close collaboration with an international network of world-renowned experts in these fields. The project is headed by a machine learning researcher PhD Markus Heinonen.