The Generative Forests

Masked autoencoder not only can be used to reconstruct images, but also can deal with videos. The only question is that how can we design such a model to achieve this purpose. It is this question, I will address in this paper. Specifically, I propose a theoretical notion - "\textbf{The Generative Spacetime}", which from the first principle, unifies the concepts of generative models in deep learning and that of Minkowski spacetime in Einstein's theory of special relativity. From the top-down physical perspective, I try to intuitively model the classical autoencoder architecture\cite{ae} as a light cone diagram to describe what we have known, by analyzing its geometric structure. Moreover, the idea of \textbf{the Generative Forests} is also derived from the geometric properties of the generative spacetime. And this work will pave the way for my future work, which is about the \textit{Hilbate Machine}.