Semantic uncertainty intervals for disentangled latent spaces

Speaker: Swami Sankaranarayanan , MIT CSAIL

Date: Wednesday, March 09, 2022

Time: 3:00 PM to 4:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: 32-D507

Event Type: Seminar

Room Description:

Host: Polina Golland, MIT CSAIL

Contact: Polina Golland, polina@csail.mit.edu

Relevant URL:

Speaker URL: None

Speaker Photo:
None

Reminders to: seminars@csail.mit.edu

Reminder Subject: TALK: Semantic uncertainty intervals for disentangled latent spaces

Providing rich and meaningful notions of uncertainty is a perennial
statistical problem which is especially widespread in computer vision,
where pixel-space uncertainty is nearly useless for decision-making.
Instead, we need uncertainty about semantically meaningful
factors---say, the age of the person in a photo, the location of a
feature, and so on. Our paper takes an important first step towards
this goal, providing rigorous uncertainty intervals on the
disentangled latent space of a generative model which has organically
extracted these semantic factors. Our procedure provides formal,
distribution-free statistical guarantees in this latent space,
regardless of the model and data used. The method does the following:
(1) it uses quantile regression to output a heuristic uncertainty
interval for each element in the latent space (2) calibrates these
uncertainties such that they provably contain the true value of the
latent for a new, unseen input with high probability. We demonstrate
that our intervals reliably communicate semantically meaningful,
statistically valid and instance-specific uncertainty in inverse
problems like image super-resolution and image completion.

Research Areas:

Impact Areas:

See other events that are part of the Biomedical Imaging and Analysis 2021 - 2022.

Created by Polina Golland at Thursday, February 24, 2022 at 2:46 PM.