MIT EECS | Quick Undergraduate Research and Innovation Scholar
Transfer Learning Across Hospitals by Discovering Latent Patient-Hospital Sub-Mechanisms
- Artificial Intelligence and Machine Learning
Hospitals with abundant patient records can train supervised machine learning models that provide effective medical prediction, but other hospitals with less data cannot do so. Transfer learnings are being developed to build models that perform well across hospitals. However, machine learning models trained using one hospital’ s data may not perform well in others due to organizational differences. Thus, we must develop methods allowing models trained in source hospitals to transfer to target hospitals. We have developed a new algorithm that combines topic modeling with a logistic regression model. Our algorithm exploits the assumption that source and target domains are mixtures of latent sub-mechanisms to learn from sourcedomain models that perform well on target domains. In this work, we implement a new type of machine learning algorithm for transfer
SuperUROP gives me an opportunity to work on a full-scale project with both technical and communication guidance, and a sense of what it’ s like to study for a doctorate. This project on machine learning in causal inference will be a good blend of knowledge I have received from previous classes in machine learning, inference, and econometrics.