Causal future prognosis in a Minkowski space-time
Estimating future events is a difficult task. Unlike humans, machine learning approaches are not regulated by a natural understanding of physics. In the wild, a plausible sequence of events is subject to the rules of causality, which cannot simply be derived from a finite training set. In this paper, researchers (Imperial College London) propose a novel theoretical framework to carry out causal predictions of the future by embedding spatiotemporal information in a Minkowski spacetime. They use the concept of the cone of light from the special theory of relativity to restrict and traverse the latent space of the anarbitrary model. They demonstrate successful applications in causal image synthesis and the prediction of future video images on an image data set. Its framework is architecture and task independent and has strong theoretical guarantees for causal capabilities.
In many everyday scenarios, we make causal predictions in order to judge how situations might develop based on our observations and experiences. Machine learning has not yet developed at this level, although automated, causally plausible predictions are highly desirable for critical applications such as medical treatment planning, autonomous vehicles, and safety. Recent work has contributed machine learning algorithms to predict the future in sequences and to causal inference. An important assumption that many approaches implicitly adopt is that the space of the model representation is a flat Euclidean space with N dimensions. However, as suggested by Arvanitidis et al. was shown, the Euclidean assumption leads to incorrect conclusions, since the latent space of a model can be better characterized as a high-dimensional, curved Riemannian space than a Euclidean space. In addition, the Alexandrov-Zeeman theorem suggests that causality requires a Lorentzian group space and advocates the unsuitability of Euclidean spaces for causal analysis. In this post, the scientists present a novel framework that changes the way we treat problems of hard computer vision such as the continuation of image sequences. They embed information in a spatio-temporal, high-dimensional, pseudo-Siemens manifold - the Minkowski spacetime - and use the special relativity concept of the light cone to carry out causal inference. You concentrate on temporal sequences and image synthesis in order to display the full capabilities of your framework.
More here