Our Team

I'm particularly interested in making AI systems safer, more reliable, and more trustworthy. The sub-field of Reliable Artificial Intelligence (RAI) I'm currently focussed on is Selective Prediction or “making models know what they don't know.” There are numerous real-life applications in which an abstention from the model is far more desirable than a wrong answer, such as Medical Imaging, Self-driving cars, visual assistants for visually impaired people, and chatbots. Unfortunately, most current models are optimized for benchmarks focusing only on the percentage of correct answers, while abstentions and wrong answers are judged equally badly. My work aims to develop new and improved abstention techniques for Multimodal Large Language Models, particularly on the Visual Question Answering (VQA) task. Let's create models that really know what they don't know!