zoom: https://umd.zoom.us/j/99147773612

Despite recent breakthroughs in aid of deep learning and scaling laws, neural networks still make a substantially fragile prediction in their worst-case behaviors, e.g., as shown in their performance degradation to adversarial examples, novelties, and corruptions. Literatures have been independently developed to address such different types of distribution shifts, they often result in bespoke models to their threat models during training, eventually limiting their applicability to recent models at scale. In this talk, I will present some of my attempts in order to make ML safety more practical and scalable, covering multiple aspects of ML safety across robustness, monitoring and alignment research. Firstly, I will introduce an Information Bottleneck based approach to mitigate the train-time dependency of robustness-aware training methods, i.e., on specific data augmentation techniques. Next, we move on our focus to adversarial robustness, and consider a crucial problem of robustness-accuracy trade-off: with an introduction of a recent technique of randomized smoothing, I will introduce a simple and scalable idea to overcome the trade-off. Lastly, I will briefly cover some of my recent works that highlights the effectiveness of language-based inference to perform efficient monitoring and uncertainty estimation.