Maggie Makar receives Google Research Scholar award for work on causally motivated AI models

Maggie Makar, assistant professor of computer science and engineering, has received a Google Research Scholar award for her project titled “Reliability through Causal Alignment: Causally Motivated Models for Pain Management.” A preeminent initiative aiming to support up-and-coming, world-class researchers, the Google Research Scholar Program provides funding for early-career professors in computer science and related fields who are performing exceptional, innovative research.

With this award, Makar proposes to improve the robustness and interpretability of machine learning models in the context of pain management by leveraging known causal relationships. The goal of her work is to provide enhanced tools that can help detect and identify pain causes and early signs of osteoarthritis, leading to better management of chronic pain and related conditions.

Artificial intelligence (AI) has emerged as a potent tool in a variety of medical applications, helping clinicians better predict, diagnose, and manage various ailments, from cancer to Covid-19. Despite their accuracy, AI models suffer when they encounter changes in data distribution, particularly due to shortcut learning, which occurs when a model learns to encode spurious correlations present in training data. This, in turn, can lead to inaccurate and biased predictions.

By using existing cause-and-effect reasoning, Makar aims to develop models that can better adapt to these distribution shifts without relying on shortcuts or requiring auxiliary data. To do this, she will use model editing techniques to identify and edit shortcuts in large language models, as well as leveraging mediators to ensure invariant learning. 

With these approaches, Makar will build machine learning models that better focus on and reflect underlying cause-and-effect relationships rather than relying on superficial correlations or shortcuts in the data. This, in turn, will result in models that are more robust to distribution shifts and less prone to biases.

Through the development of more robust, generalizable models built around causal mechanisms, Makar is providing a crucial tool for ensuring accurate diagnoses, effective treatments, and equitable outcomes in the context of chronic pain.