Robust machine learning for reliable deployment
Wednesday, Oct 13, 2021, 10:00am-11:00pm (CEST)
I will give an overview of our recent advances in reliable machine learning, including weakly supervised learning, learning from noisy supervision, and learning under distribution shift. Specifically, for learning from weak supervision, we introduce an empirical risk minimization framework that allows estimation of the classification risk only from weakly supervised data (such as positive-unlabeled data, positive-confidence data, similar-dissimilar data, complementary labels, and partial labels). This framework is fairly general and is compatible with any loss function, classification model, and optimizer.
Next, for learning from noisy supervision, we provide loss correction methods based on noise transition estimation and sample selection methods based on the memorization effect of deep learning. These methods can overcome the limitations of standard noise-robust approaches based on robust statistics and regularization.
Finally, for distribution shift, we present a novel covariate shift adaptation method that allows simultaneous learning of the importance weight and predictor and a dynamic importance weighting method that allows mini-batch-wise adaptation for general distribution shift.
Masashi Sugiyama received the PhD degree in Computer Science from Tokyo Institute of Technology, Japan, in 2001. After experiencing
assistant and associate professors at the same institute, he became a professor at the University of Tokyo in 2014.
Since 2016, he has concurrently served as Director of RIKEN Center for Advanced Intelligence Project. He (co-)authored machine learning monographs including Machine Learning in Non-Stationary Environments (MIT Press, 2012), Density Ratio Estimation in Machine Learning (Cambridge University Press, 2012), Statistical Reinforcement Learning (Chapman and Hall, 2015), Introduction to Statistical Machine Learning (Morgan Kaufmann, 2015), Variational Bayesian Learning Theory (Cambridge University Press, 2019), and Machine Learning from Weak Supervision (MIT Press, in press). He served as a program co-chair for NeurIPS2015 and AISTATS2019.