EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0756
Title: Learning the data manifold for reusable augmentations Authors:  Kion Fallah - Georgia Institute of Technology (United States) [presenting]
Marissa Connor - Embedded Intelligence (United States)
Christopher Rozell - Georgia Institute of Technology (United States)
Abstract: The manifold hypothesis suggests that variations in high-dimensional, real-world data lie on or near a low-dimensional manifold. We discuss recent work to learn this manifold from data by incorporating a generative manifold model in the latent space of a deep auto-encoder. This model represents the manifold with a dictionary of Lie group operators, representing the non-linear path between any two data points with a sparse combination of dictionary entries. To speed up training, we propose an inference procedure that can be quickly run on a GPU. After unsupervised training, we demonstrate that the learned Lie group operators are re-usable across a dataset for generating semantically meaningful augmentations.