Project description
New algorithms for machine learning of physical behavior in space and time
Numerical simulations are frequently required to study the behaviour of systems whose mathematical models are too complex to provide analytical solutions. In studying these phenomena, it is necessary to consider spatial dimensions that are difficult to compute and store in terms of the resources required to do so. The EU-funded SpaTe project aims to develop novel algorithms to infer spatio-temporal functions, allowing also for the construction of efficient representations that developers say will tame their complexity and high dimensionality. Ultimately, SpaTe will allow for a better understanding of the physical world around us and offer substantial practical applications spanning the range from social media apps to self-driving cars.
Objective
Numerical simulations are of tremendous importance for a wide range of scientific disciplines and commercial enterprises. For the majority of these natural phenomena, we not only need to consider the three spatial dimensions, but additionally we need to resolve how these phenomena develop over time. Thus, most natural simulations inherently need to resolve four dimensional functions, and most effects at human scales require fine discretizations along all four axes. As a consequence, these functions require large amounts of resources to compute and store. This problem becomes even more pronounced with the advent of data-driven techniques and machine learning. The learning algorithms effectively add additional dimensions, and the complexity and dimensionality of the corresponding functions explains the current lack of data-driven algorithms for space-time functions despite their enormous potential. Within this research project I plan to address the fundamental difficulties that arise in this setting: I will develop novel algorithms to infer spatio-temporal functions, and to construct efficient representations to tame their complexity and high dimensionality. This project combines numerical simulations with computer vision, and machine learning, and has the potential to radically change the way we work with physical simulations. Not only will it break new ground for fast and controllable VFX animations, but it will additionally facilitate the development of new ways to capture physical effects, in conjunction with algorithms to make physical predictions based on observations. Ultimately, this direction will allow us to better understand the physical world around us. It will help us to analyze sparse and ambiguous measurements such as videos and 3D scans automatically and reliably, with a vast range of practical applications from social-media apps to autonomous vehicles.
Fields of science
Programme(s)
Funding Scheme
ERC-COG - Consolidator GrantHost institution
80333 Muenchen
Germany