Speaker
Edmund Ross
Description
This paper presents theoretical and empirical results demonstrating the effectiveness of "Tokenization" in capturing the physical operator of an ergodic flow, a technique used in multiple long video machine learning architectures, in particular in NVIDIA's LongVideoGAN. Complementing the theory, we present novel experimental results on canonical problems, ranging from the wave and heat equations to the Kuramoto-Sivashinsky (KS) equation, using both LongVideoGAN and autoregressive models. Finally, we show LongVideoGAN is able to learn a computational fluid dynamics (CFD) dataset featuring a Kármán vortex street, with excellent temporal correlation and generalisation results.
