Be Confident in What You Know: Bayesian Parameter Efficient Fine-Tuning of Vision Foundation Models

Exhibit will be through poster presentation of the research work done in the artificial intelligence/machine learning. Visitors will learn about the issues that come with few shot adaptation of foundation models, and how to tackle them. Few-shot adaptation with vision foundation models achieve high accuracy, but model is highly under-confident while predicting the output. We leverage evidential theory to explain the under-confidence issue and propose novel formulation to tackle the underconfidence issue.