Artificial Intelligence Master's students Luca Falorsi, Pim de Haan and Tim Davidson will give a plenary talk on 'Reparameterizing distributions' at the leading Machine Learning conference AISTATS, held between 16-18 April in Japan. The students developed a method to describe the movement of static objects in 3D.
The so-called reparameterization trick, in part developed at the UvA, is an important tool used in (unsupervised) machine learning, to directly learn distributions for observed data. It describes a way of sampling from a distribution, in a way that makes it possible to optimise the parameters of said distribution directly using a (deep) machine learning architecture.
While vital for many modern approaches, the trick unfortunately is not yet available for all distributions - notably, it is not known how to construct one for most distributions defined on spaces with a non-trivial topology. This is problematic, as many results in fields such as Physics and Robotics, make extensive use of such spaces to, for example, model the 3D movements of rigid body objects.
In the work the students will present at the conference, 'Reparameterizing Distributions on Lie Groups', they demonstrate a novel way to extend this trick to a large class of these non-trivial spaces called ‘Lie Groups’. Using their extension, it now becomes possible to realistically model complex system dynamics using a variety of distributions previously unavailable. The Master's students developed the methodology under supervision of Patrick Forré, postdoc at the Informatics Institute.
The International Conference on Artificial Intelligence and Statistics, or simply AISTATS, is a leading AI conference . With the ever growing popularity of AI and Machine Learning, this year there were 1,111 submissions, of which only 28 were invited to present; the UvA students being among them.