Are you passionate about bleeding edge research on
Deep Learning and Dynamical Systems, (Stochastic) Partial Differential Equations, for Sciences? Can we discover, model and explain dynamics inside Neural Networks? Can we use Neural Networks to solve better, more efficiently, more generally, Ordinary and Partial Differential Equations, or stochastic versions of them?
We are looking for a postdoctoral researcher with either Machine Learning, Physics, Applied Mathematics, or Computer Vision background to join a team of 20 researchers working on these very topics (2 years contract); a team that is connected with the
ELLIS Network of Excellence in AI; a team with consistent and strong presence in the top Machine Learning and Computer Vision conferences and journals.
What are you going to doBeyond static data, static learning, static algorithms,
the next innovation in Machine and Deep Learning will be in
the direction of Neural Networks Dynamics, Dynamical Systems, and PDEs with focus on scientific data. We are looking for a Postdoctoral researcher who is just as passionate to invest in this direction.
Understanding space and time in machine learning, be it with neural networks, Markov models, deep probabilistic models, and beyond, is one of the biggest problems, especially given the vast availability of video and other high-dimensional time series data. While consumer videos are unconstrained and thus unclear how or what to model about them, 'videos' of spatiotemporal scientific recording offer a great opportunity for innovation. For one, they pertain certain and complex spatiotemporal dynamics that modern machine learning algorithms cannot easily model beyond restricted cases. What is more, the underlying science of these dynamics can be used to evaluate and compare algorithms rather than relying on biased human annotations for that. Third, scientific data are almost always spatiotemporal, they are vast and grow at an immense rate, and are ready to be understood by algorithms.
Given all the evidence and potential impact, part of this position will also be co-designing and creating a Spatiotemporal Dynamical Systems Decathlon, which would hopefully act as a catalyst for future research in this direction. This is an ongoing initiative already started and led by the PI, E. Gavves. In this Decathlon we will be organizing scientific data from different disciplines together with respective leading experts, and designing experiments that showcase the value of learning to simulate physical processes across fields.
Learning to simulate is and will keep being an exciting and entirely novel direction of machine learning research both in academia and industry. The PostDoctoral researcher will have the opportunity to work with a team of 20 Ph.D. students working on these problems. The funding is from personal grants with little strings attached, and
fundamental research is possible and desirable.Tasks and responsibilities:
- show independence in achieving research goals and willingness to collaborate and to supervise PhD students working on deep learning and dynamical systems, differential geometry, sciences;
- co-lead a large initiative of researchers from multiple disciplines (physics, fluid dynamics, astronomy, biology, chemistry) on developing the Spatiotemporal Dynamical Systems Decathlon;
- invent, evaluate, and describe novel learning algorithms for spatiotemporal PDE and dynamical systems;
- contribute to a real-world showcase demo;
- present research results at international conferences, workshops, and journals;
- become an active member of the research community and to collaborate with other researchers, both within and outside the Informatics Institute;
- contribute to teaching activities, such as lectures, lab courses, or supervising bachelor and master students.