Uncertainty and Robustness in Deep Visual Learning

News

Poster spotlight presentations are now available.

Accepted papers are now available at CVF open access archive.

About

The past decade was marked by significant progress in the field of artificial intelligence and statistical learning. Efficient new algorithms, coupled with the availability of large datasets and dramatic increase in computing power, led to solutions that match, or exceed, human performance in perception tasks such as image and speech recognition, as well as in building strategies for complex board and computer games. Deep learning technologies are being widely applied across different areas, from new art form creation to subatomic particle and drug discovery. However, the most efficient of modern models come in form of a black box, with the majority of them lacking the ability to robustly reason about the confidence of their predictions. At the same time, being capable to quantify model uncertainty and recognize failure scenarios is crucial when it comes to incorporating them into complex decision making pipelines, e.g. autonomous driving or medical image analysis systems.

This workshop will bring together researchers focused on the discovery of new architectures and statistical methods for robust predictive models, as well as computer vision and machine learning practitioners from the industry that work on applying those models in complex real-world scenarios. The goal of the workshop is to build a solid understanding of how classical uncertainty quantification methods could be applied in the deep learning era. It will also provide the discussion of methods developed over the past decade and solutions for efficient industrial applications.

Covered topics

• Uncertainty quantification;
• Deep probabilistic models;
• Failure prediction;
• Bayesian methods;
• Robust vision in autonomous driving, robotics, medical imaging;
• Open-world, lifelong learning;
• Uncertainty quantification in domain transfer, active learning, adversarial attack defense.

Workshop Schedule

Location: Hyatt Beacon A (Long Beach Convention Center)

Poster session location: Pacific Arena Ballroom, posters 138-161.

13:40 Introduction

13:45 Invited talk: Sergey Levine (UC Berkeley)

14:15 Poster Spotlight Session

• 14:15 "Improving Deep Network Robustness to Unknown Inputs with Objectosphere"
• 14:20 "Incremental Learning with Unlabeled Data in the Wild"
• 14:25 "Iterative Self-Learning: Semi-Supervised Improvement to Dataset Volumes and Model Accuracy"
• 14:30 "Measuring Calibration in Deep Learning"
• 14:35 "On the Sensitivity of Adversarial Robustness to Input Data Distributions"

14:40 Invited Talk: Andreas Geiger (Max Planck Institute, University of Tübingen)

15:10 Poster session/Coffee break. Location: Pacific Arena Ballroom. Poster locations by paper title.

16:30 Invited Talk: Kris Kitani (Carnegie Mellon University)

17:00 Invited Talk: Alex Kendall (Wayve, University of Cambridge)

17:30 Invited Talk: Raquel Urtasun (Uber, University of Toronto)

18:00 Workshop closes

Invited speakers

Call for papers

We invite submission of the extended abstracts (using the CVPR 2019 format) describing work in the domains suggested above or in closely-related areas. Accepted papers will be presented during the spotlight\poster session at the workshop and appear in CVF open access archive. Extended abstract size limit is 3 pages without references. Authors may submit a draft of their poster as an optional fourth page. The review process is double-blind.

Submissions of work which have been previously published, including papers accepted to the main CVPR 2019 conference, are allowed.

Paper submission deadline: April 15, 2019

Author notification: May 6, 2019

Camera-ready deadline: May 13, 2019

Organizers

Sponsors