Search
Workshop

Geometric disentanglement by random convex polytopes

  • Michael Joswig (TU Berlin & MPI MiS, Berlin, Germany)
Live Stream MPI für Mathematik in den Naturwissenschaften Leipzig (Live Stream)

Abstract

We propose a new geometric method for measuring the quality of representations obtained from deep learning. Our approach, called "Random Polytope Descriptor", provides an efficient description of data points based on the construction of random convex polytopes. We demonstrate the use of our technique by qualitatively comparing the behavior of classic and regularized autoencoders. This reveals that applying regularization to autoencoder networks may decrease the out-of-distribution detection performance in latent space. While our technique is similar in spirit to k-means clustering, we achieve significantly better false positive/negative balance in clustering tasks on autoencoded datasets.

Joint work with Marek Kaluba and Lukas Ruff.

Links

conference
4/6/21 4/9/21

(Polytop)ics: Recent advances on polytopes

MPI für Mathematik in den Naturwissenschaften Leipzig Live Stream

Saskia Gutzschebauch

Max Planck Institute for Mathematics in the Sciences Contact via Mail

Federico Castillo

Max Planck Institute for Mathematics in the Sciences

Giulia Codenotti

Goethe University Frankfurt

Benjamin Schröter

Royal Institute of Technology (KTH)