Timezone: »

SelecMix: Debiased Learning by Contradicting-pair Sampling
Inwoo Hwang · Sangjun Lee · Yunhyeok Kwak · Seong Joon Oh · Damien Teney · Jin-Hwa Kim · Byoung-Tak Zhang

Tue Nov 29 02:00 PM -- 04:00 PM (PST) @ Hall J #426

Neural networks trained with ERM (empirical risk minimization) sometimes learn unintended decision rules, in particular when their training data is biased, i.e., when training labels are strongly correlated with undesirable features. To prevent a network from learning such features, recent methods augment training data such that examples displaying spurious correlations (i.e., bias-aligned examples) become a minority, whereas the other, bias-conflicting examples become prevalent. However, these approaches are sometimes difficult to train and scale to real-world data because they rely on generative models or disentangled representations. We propose an alternative based on mixup, a popular augmentation that creates convex combinations of training examples. Our method, coined SelecMix, applies mixup to contradicting pairs of examples, defined as showing either (i) the same label but dissimilar biased features, or (ii) different labels but similar biased features. Identifying such pairs requires comparing examples with respect to unknown biased features. For this, we utilize an auxiliary contrastive model with the popular heuristic that biased features are learned preferentially during training. Experiments on standard benchmarks demonstrate the effectiveness of the method, in particular when label noise complicates the identification of bias-conflicting examples.

Author Information

Inwoo Hwang (Seoul National University)
Sangjun Lee (Seoul National University)
Yunhyeok Kwak (Seoul National University)
Seong Joon Oh (University of Tübingen)
Damien Teney (Idiap Research Institute)
Jin-Hwa Kim (NAVER AI Lab)
Jin-Hwa Kim

Jin-Hwa Kim has been Technical Leader and Research Scientist at NAVER AI Lab since August 2021 and Guest Assistant Professor at Artificial Intelligence Institute of Seoul National University (SNU AIIS) since August 2022. He has been studying multimodal deep learning (e.g., [visual question answering](http://visualqa.org)), multimodal generation, ethical AI, and other related topics. In 2018, he received Ph.D. from Seoul National University under the supervision of Professor [Byoung-Tak Zhang](https://bi.snu.ac.kr/~btzhang/) for the work on "Multimodal Deep Learning for Visually-grounded Reasoning." In September 2017, he received [2017 Google Ph.D. Fellowship](https://ai.googleblog.com/2017/09/highlights-from-annual-google-phd.html) in Machine Learning, Ph.D. Completion Scholarship by Seoul National University, and the VQA Challenge 2018 runners-up at the [CVPR 2018 VQA Challenge and Visual Dialog Workshop](https://visualqa.org/workshop_2018.html). He was Research Intern at [Facebook AI Research](https://research.fb.com/category/facebook-ai-research/) (Menlo Park, CA) mentored by [Yuandong Tian](http://yuandong-tian.com), [Devi Parikh](https://www.cc.gatech.edu/~parikh/), and [Dhruv Batra](https://www.cc.gatech.edu/~dbatra/), from January to May in 2017. He had worked for SK Telecom (August 2018 to July 2021) and SK Communications (January 2011 to October 2012).

Byoung-Tak Zhang (Seoul National University)

More from the Same Authors