Srijita Das

By |

My research is about building sample-efficient machine learning models. My long term goal is to develop collaborative systems that can actively seek advice from humans and make faster decisions, resulting in reliable and practical systems. I specifically focus on design of sequential decision-making models to make them learn faster. We leverage advice from humans in various forms (implicit and explicit) to encourage favorable decisions and avoid decisions having catastrophic consequences. We also focus on minimizing the cost of seeking advice by building suitable machine learning models from historical advice data and reusing them when required. Our research also develops ways to solve complex tasks in Reinforcement Learning by leveraging various kinds of knowledge transfer mechanisms, curriculum learning, teacher-student framework etc. Advances in these directions would make decision-making models sample-efficient and better suited for solving real-world problems. Along the supervised machine learning spectrum, we also focus on problems related to learning with less data, traditionally known as Active Learning, semi-supervised learning, and learning from multiple experts.