Dandelion

In the the Data Curation for Trustworthy AI project (report can be found here) we define bias as when a model performs unequally along the dimension of a selected characteristic in the data for which the user expects. Dandelion aims to first be able to measure the amount of bias present in any given multimodal foundation model as defined by a dynamic selected characteristic provided by the user at runtime. Then Dandelion will identify the dimension and location of that bias in the model itself, and further attempt to erase the concept of the selected characteristic from the model.

This project is in progress. I am the PI on this project for a small team. This project was an internally funded research project that I pitched and won a year-long grant for.