Impact of limited and noisy data on trustworthy machine learning

Machine Learning (ML) algorithms are known to suffer from various issues when it comes to their trustworthiness including properties like adversarial robustness, robustness to distribution shift, privacy, and fairness. Responsibly deploying ML algorithms in critical applications requires satisfying these notions of trustworthiness. While there has been significant progress in developing new trustworthy algorithms, the role of inadequate data is often ignored. Widely available data in the real world are often noisy, limited, and long-tailed and play a role in hindering these aspects of trustworthiness. In this talk, we will look at characterising some of the fundamental limitations on trustworthiness due to inadequate data. In the second part of the talk, we will look at overcoming some of these limitations with plausible relaxations and new algorithms. Finally, we will conclude with potential future directions in this space.