Reproduce “All Centers Are at most a Few Tokens Apart: Knowledge Distillation with Domain Invariant Prompt Tuning” on OpenResearch

About this paper

Amir Mohammad Ezzati, Alireza Malekhosseini, Armin Khosravi, Mohammad Hossein Rohban

A method called Domain Invariant Prompt Tuning (DIPT) was developed at Sharif University of Technology to improve domain generalization in computational pathology by learning domain-specific continuous prompts. This approach, integrated into a knowledge distillation pipeline, increased mean F1 scores by up to 3.08% and worst-case F1 scores by up to 6.05% on the Camelyon17-WILDS dataset.

Key insights

• Implicitly learning domain characteristics through data-driven prompt tuning effectively mitigates domain shift in computational pathology, even without explicit semantic descriptors.
• Aggregating domain-specific learned prompts into domain-invariant class-generic embeddings provides a potent source of generalizable knowledge for student models during distillation.
• The novel integration of prompt tuning, Vision-Language Models (VLMs), and knowledge distillation enhances model robustness and generalization crucial for practical AI adoption in pathology.

Summary by alphaXiv.