New architectures for reliable inference and predictions in high-dimensional multimodal health and biomedicine data
We are developing AI models for clinical and biomedical data for meaningful downstream tasks. Brute forcing existing inductive biases and architectures can lead to unreliable scientific inference and predictions. Therefore, the lab focuses on developing data representations, learning tasks, and model architectures that will be robust to large distributional shifts, and evaluated with robust frameworks to test inferential capabilities.
Adaptive AI systems and representations to improve generalizability and robustness
Enabling systems that generalize to out-of-distribution data requires adaptive AI approaches where models can improve from their mistakes. We leverage probabilistic modeling, reinforcement learning, and deep learning to overcome imperfections of observational health and medicine data and develop adaptive multimodal AI systems.
Daksh Mittal, Yuanzhe Ma, Shalmali Joshi, and Hongseok Namkoong.
Neural Information Processing Systems (NeurIPS) 2024
New computational methods for specific applications in health and medicine
Our work involves models being developed in-house as well as in collaboration with other clinical collaborators and healthcare institutions. These insights help us inform the gaps in current shortcomings of foundation models and in our ability to use AI as expert reasoning agents. Our current applications are in psychiatry, cardiology, radiology, and neurocritical care.
Haoran Zhang, Harvineet Singh, Marzyeh Ghassemi, and Shalmali Joshi
International Conference on Machine Learning (ICML) 2023