Timezone: »

 
Poster
Improving Robustness against Real-World and Worst-Case Distribution Shifts through Decision Region Quantification
Leo Schwinn · Leon Bungert · An Nguyen · René Raab · Falk Pulsmeyer · Doina Precup · Bjoern Eskofier · Dario Zanca

Wed Jul 20 03:30 PM -- 05:30 PM (PDT) @ #428
in Poster Session 2 »

The reliability of neural networks is essential for their use in safety-critical applications. Existing approaches generally aim at improving the robustness of neural networks to either real-world distribution shifts (e.g., common corruptions and perturbations, spatial transformations, and natural adversarial examples) or worst-case distribution shifts (e.g., optimized adversarial examples). In this work, we propose the Decision Region Quantification (DRQ) algorithm to improve the robustness of any differentiable pre-trained model against both real-world and worst-case distribution shifts in the data. DRQ analyzes the robustness of local decision regions in the vicinity of a given data point to make more reliable predictions. We theoretically motivate the DRQ algorithm by showing that it effectively smooths spurious local extrema in the decision surface. Furthermore, we propose an implementation using targeted and untargeted adversarial attacks. An extensive empirical evaluation shows that DRQ increases the robustness of adversarially and non-adversarially trained models against real-world and worst-case distribution shifts on several computer vision benchmark datasets.

Keywords: DL: Algorithms · DL: Robustness
[ Slides]  [ Poster]  [ Paper PDF] 

Author Information

Leo Schwinn (Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU))
Leon Bungert (University of Bonn)
An Nguyen (Friedrich-Alexander-Universität Erlangen-Nürnberg)
René Raab (Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU))
Falk Pulsmeyer (Friedrich-Alexander University Erlangen-Nürnberg)
Doina Precup (McGill University / DeepMind)
Bjoern Eskofier (Friedrich-Alexander-Universität Erlangen-Nürnberg)
Bjoern Eskofier

Bjoern M. Eskofier is German Research Foundation (DFG) funded Heisenberg-Professor for "Digital Support Systems in Sports and Medical Engineering". He heads the Machine Learning and Data Analytics (MaD) Lab at the Friedrich-Alexander-University Erlangen-Nuernberg (FAU). He is also the current speaker of FAU’s Department Artificial Intelligence in Biomedical Engineering (AIBE) and the co-speaker of the German Research Foundation collaborative research center “EmpkinS” (SFB 1483). Currently, the MaD Lab has 45 co-workers, who research in the fields of machine learning and artificial intelligence mainly for ubiquitous computing systems in “digital” healthcare and sports. The motivation of the lab’s researchers is to increase human wellbeing. Dr. Eskofier studied Electrical Engineering at the FAU and graduated in 2006. He then studied under the supervision of Prof. Dr. Benno Nigg at the University of Calgary (Canada). There, he received his PhD degree in Biomechanics in 2010 for his research on "Application of Pattern Recognition Methods in Biomechanics". He authored more than 260 peer reviewed articles, submitted 5 patent applications, and started three spinoff startup companies. He won several medical-technical research awards, including the “Curious Minds” award in the category “Life Sciences” given by Manager Magazin and Merck. In 2016, he was a visiting professor in Dr. Paolo Bonato’s Motion Analysis Lab at Harvard Medical School (February-March), and in 2018, he was a visiting professor in Dr. Alex “Sandy” Pentland’s Human Dynamics group at MIT Media Lab (March-August). He is also a delegate of the FAU to the Medical Valley (80 Mio Euro German Ministry of Education funded cluster) and to the European Institute of Innovation & Technology for Health (EIT Health, 500 Mio Euro EU consortium, 2015-2025). Bjoern Eskofier has defined his research and entrepreneurial agenda to revolve around contributions to a “Digital Health Ecosystem”, where patients are connected to other stakeholders within the Healthcare system using digital support tools. His digital health research philosophy is that only multidisciplinary teams of engineers, medical experts, industry representatives and entrepreneurs will have the tools to actually implement changes in Healthcare.

Dario Zanca (Friedrich-Alexander-Universität Erlangen-Nürnberg)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors