Federated Learning of Plug-and-Play Adapter for Segment Anything Model
Date of Award
4-30-2024
Document Type
Thesis
Degree Name
Master of Science in Computer Vision
Department
Computer Vision
First Advisor
Prof. Mohammad Yaqub
Second Advisor
Prof. Karthik Nandakumar
Abstract
Foundation models trained on natural images exhibit strong generalization capabilities, requiring only minimal fine-tuning across various downstream tasks. However, adapting these models for medical image analysis is challenging due to extreme distribution shifts compared to pre-trained source data. This is further exacerbated by privacy constraints that inhibit siloed task-specific medical data pooling at a central location for accurate fine-tuning. This work addresses this challenge by leveraging the combined strengths of Parameter-Efficient Fine-tuning (PEFT) and Federated learning (FL). Specifically, we learn plug-and-play Low-Rank Adapters (LoRA) in a federated manner to adapt the Segment Anything Model (SAM) for 3D medical image segmentation without modifying any parameters of the original SAM model. Our experiments show that retaining parameters in their original state during adaptation is beneficial because fine-tuning them tends to distort the inherent capabilities of the underlying foundation model. Furthermore, PEFT complements FL by decreasing communication cost (∼49× ↓) compared to full fine-tuning (FullFT), while also substantially outperforming FullFT (∼6% ↑ Dice score) in 3D segmentation tasks on Fed-KiTS19 dataset.
Recommended Citation
M. Asokan, "Federated Learning of Plug-and-Play Adapter for Segment Anything Model,", Apr 2024.
Comments
Thesis submitted to the Deanship of Graduate and Postdoctoral Studies
In partial fulfilment of the requirements for the M.Sc degree in Computer Vision
Advisors: Mohammad Yaqub, Karthik Nandakumar
with 2 years embargo period