Generalization capability to unseen domains is crucial for machine learning models when deploying to real-world conditions. We investigate the challenging problem of domain generalization, i.e., training a model on multi-domain source data such that it can directly generalize to target domains with unknown statistics. We adopt a model-agnostic learning paradigm with gradient-based meta-train and meta-test procedures to expose the optimization to domain shift. Further, we introduce two complementary losses which explicitly regularize the semantic structure of the feature space. Globally, we align a derived soft confusion matrix to preserve general knowledge of inter-class relationships. Locally, we promote domain-independent class-specific cohesion and separation of sample features with a metric-learning component. The effectiveness of our method is demonstrated with new state-of-the-art results on two common object recognition benchmarks. Our method also shows consistent improvement on a medical image segmentation task.
Qi Dou (Imperial College London)
Dr. Qi DOU is a postdoctoral research associate at the Department of Computing at Imperial College London. Before that, she has received her Ph.D. degree in Computer Science and Engineering at The Chinese University of Hong Kong in July 2018. Her research interests are in the development of advanced machine learning methods for healthcare applications with specifics to medical image computing. Dr. Dou has won the Best Paper Award of Medical Image Analysis-MICCAI in 2017, the Best Paper Award of Medical Imaging and Augmented Reality in 2016, and MICCAI Young Scientist Award Runner-up in 2016. She has also won the HKIS Young Scientist Award 2018. She is going to join CUHK as an Assistant Professor in Jan 2020.
Daniel Coelho de Castro (Imperial College London)
I am a final-year PhD student in the Biomedical Image Analysis group at Imperial College London, supervised by Dr Ben Glocker. My BSc in Computer Engineering was obtained from the Pontifical Catholic University of Rio de Janeiro, followed by a graduate Engineering Diploma from École Centrale Paris (now CentraleSupélec) and a MRes in Advanced Computing from Imperial College. I have additionally done a long research internship at Microsoft Research Cambridge, under supervision of Dr Sebastian Nowozin. My research interests revolve around probabilistic modelling of imaging and non-imaging data, especially involving (but not limited to) Bayesian nonparametrics, deep generative models, representation learning and disentanglement, causality, and shape analysis.
Konstantinos Kamnitsas (Imperial College London)
Ben Glocker (Imperial College London)
More from the Same Authors
2019 Workshop: Medical Imaging meets NeurIPS »
Hervé Lombaert · Ben Glocker · Ender Konukoglu · Marleen de Bruijne · Aasa Feragen · Ipek Oguz · Jonas Teuwen
2018 Workshop: Medical Imaging meets NIPS »
Ender Konukoglu · Ben Glocker · Hervé Lombaert · Marleen de Bruijne
2017 Workshop: Medical Imaging meets NIPS »
Ben Glocker · Ender Konukoglu · Hervé Lombaert · Kanwal Bhatia