Certifiable Out-of-Distribution Generalization

• DOI: 10.1609/aaai.v37i9.26295
• Publication Date: 2023-06-27T17:51:59Z
• Authors: Nanyang Ye; Lin Zhu; Jia Wang; Zhaoyu Zeng; Jiayao Shao; Chensheng Peng; Bikang Pan; Kaican Li; Jun Zhu
• Keywords: Margin (machine learning); Code (set theory)

Machine learning methods suffer from test-time performance degeneration when faced with out-of-distribution (OoD) data whose distribution is not necessarily the same as training distribution. Although a plethora of algorithms have been proposed to mitigate this issue, it has demonstrated that achieving better than ERM simultaneously on different types distributional shift datasets challenging for existing approaches. Besides, unknown how and what extent these work any OoD datum without theoretical guarantees. In paper, we propose certifiable generalization method provides provable guarantees via functional optimization framework leveraging random distributions max-margin each input datum. With approach, algorithmic scheme can provide certified accuracy datum's prediction semantic space achieves dominated by correlation shifts or diversity shifts. Our code available at https://github.com/ZlatanWilliams/StochasticDisturbanceLearning.