ROBOSHOT: Improving Foundation Model Performance Without Fine-Tuning

Discover how to enhance foundation model performance without fine-tuning in this 36-minute research presentation by PhD Student Dyah Adila from the University of Wisconsin-Madison. Learn about the ROBOSHOT method, which improves the robustness of zero-shot embeddings by leveraging large language models to identify helpful and distracting features. Explore techniques to boost pretrained model performance, reduce harmful biases, and understand the conditions for optimal results. Gain insights into improving the robustness of pretrained model embeddings in a fully zero-shot fashion and characterizing performance enhancement conditions.

ROBOSHOT: better foundation model performance without fine-tuning (Stanford researcher presentation)