| Abstract: |
| In machine learning, bounded loss functions have been more and more frequently used owing to their robustness to outliers and heavy-tailed noise. However, the understanding of bounded loss functions, especially from a theoretical viewpoint, is still limited due to their nonconvexity. In this talk, I will report some of our recent efforts made in this regard. First, I will show that in the context of empirical risk minimization, bounded loss functions can be interpreted from a minimum distance estimation viewpoint. Second, the prediction ability of estimators resulting from bounded loss functions will also be assessed and discussed. |
|