Dimitris Tsipras (Standford University)
Dimitris Tsipras is a postdoc at Stanford University advised by Percy Liang and Greg Valiant. He obtained his PhD from MIT where he was advised by Aleksander Madry. His research is focused on understanding and improving the reliability of modern machine learning methods.
Short Abstract:The canonical recipe for building a machine learning system involves training a model on a dataset that is designed to capture a specific task. However, creating a good dataset is not always easy. In this talk, I will describe how different design choices can cause our dataset to systematically deviate from the real-world task it is intended to capture. I will then present a new paradigm for interacting with models after training by rewriting their predictions rules. This paradigm allows a model designer to directly modify the behavior of the model to adapt it to new environments or fix bugs in its behavior.