Unsupervised learning of geometrical features from images by explicit group actions enforcement

Luca Bottero (1), Francesco Calisto (2), Valerio Pagliarino (3) - September 2022

(1) Università degli Studi di Torino (student), Machine Learning Journal Club
(2) Ludwig-Maximilians-Universität München and Technische Universität München (student), Machine Learning Journal Club
(3) Università degli Studi di Torino (student), Machine Learning Journal Club

Proposal for NeurReps Workshop (NeurIPS conference), Symmetry and Geometry in Neural Representations. New Orleans, 3rd December 2022.

Abstract

In this work we propose an autoencoder architecture capable of automatically learning meaningful geometric features of objects in images, achieving a disentangled representation of 2D objects. It is made of a standard dense autoencoder that captures the \textit{deep features} identifying the shapes, and an additional encoder that extracts geometric latent variables regressed in an unsupervised manner, that are then used to apply a transformation on the output of the \textit{deep features} decoder. We show promising results and that this approach performs better than a non-constrained model having more degrees of freedom.

Keywords:
Autoencoders, group actions, geometric priors, latent space disentanglement

Dependencies

importlib-metadata  4.12.0
importlib-resources 5.2.2
h5py                2.10.0
matplotlib          3.4.3
numpy               1.19.5
progressbar         2.5
pytorch-ranger      0.1.1
seaborn             0.11.2
torch               1.9.1
torch-optimizer     0.1.0
torchsummary        1.5.1
torchvision         0.10.1
tqdm                4.62.2