Efficient Adaptation of Trained Models When Utilities of Model Predictions Change
PI: Vasant Honavar (Informatics and Intelligent Systems)
Utility-based learning (UBL) is concerned with building ML systems which optimize applicationspecific objectives, expressed using utility functions that model the differing utilities of model predictions. For example, when fighting a pandemic, the health system may assign greater utility to reliably identifying those at risk of the disease so they can be vaccinated as compared to identifying those at risk of developing a chronic illness such as diabetes for targeting with lifestyle change interventions. In practical applications, it is inevitable that the utility of different model predictions changes over time. This raises the question as to if, how, and under what conditions a trained model can be rapidly adapted (without necessitating complete retraining) when faced with changes in utilities of predictions. While utility based machine learning has been explored in settings where the utility function is pre-specified and constant, there is little work on settings where the utility function changes after the trained model is deployed.
Objectives
This project has the following aims:
• Precise mathematical formulation of the problem of adapting trained models in the face of changes to utilities of model predictions
• Characterizations of the conditions under which a trained model can be efficiently adapted when the utilities of model predictions change
• Design of algorithms with provable performance guarantees relative to the best (complete retraining of the model) and the worst case scenarios (no model adaptation in response to change in utilities)
• Application of the resulting algorithms to applications such as healthcare, disaster relief, etc.
Long-term goal
The long-term goal of this project is to develop methods for efficient adaptation of predictive models trained using machine learning when the utilities of model predictions change, with practical applications across a broad range of real-world applications
Connection with ICDS Mission
This work directly contributes to the development of new methods that would extend the practical applicability and effectiveness of machine learning in settings where the utilities of model predictions change after the model is deployed e.g., healthcare, disaster relief, etc.