The Machine Learning of Time and Dynamics … with an Outlook towards the Sciences
By Dr. Efstratios Gavves (University of Amsterdam)
In the past decades, the impressive progress in machine learning and applications -like computer vision- was mainly by assuming (or enforcing) that data is static and usually of spatial-only nature, that data is i.i.d, that learning correlations suffices for high predictive accuracies. In the real world, however, data and processes are typically (spatio-) temporal, dynamic, non-stationary, non-iid, causal. This leads to paradoxical situations for learning algorithms. In this talk, I will first present my vision for a new type of learning that embraces temporality and dynamics. I will then discuss recent work that connects complexity in deep stochastic models, like hierarchical VAEs, with phase transitions, pointing perhaps to a link to statistical physics. I will continue with discussing how simple ways of introducing roto-translation equivariance can greatly improve standard neural relational inference in modelling dynamics of complex interacting dynamical systems. Last, I will present our latest attempts in scaling up causal discovery by at least two orders of magnitude compared to the recent literature. I will close with drawing a connection between machine learning and the sciences, whose interface -I believe- is deeply temporal and dynamical, and will inspire the great next breakthroughs.
We have three more speakers at this event:
• Bart ter Haar Romenij (TU/e), our society president will give his farewell lecture on something quite special: understanding deep learning from first principles and the human visual system. To support the credibility: all will be explained with life coding.
• Alex Frangi (Leeds Univ.), an outstanding computer vision researcher from England tells about research at the crossroads of image analysis and modeling with emphasis on machine learning.
• Veronika Cheplygina(IT Univ. of Copenhagen), also leaving the board, has an exciting lecture about shortcomings of machine learning and recommendations to overcome these.
Share this Post