Transferability Score

Transferability scores aim to predict how well a pre-trained model will perform on a new, downstream task without the computationally expensive process of fine-tuning it on every potential model. Current research focuses on developing and benchmarking these scores across diverse domains, including image classification, object detection, medical imaging, and speech processing, employing various techniques like Bayesian likelihood estimation and optimal transport. While promising results exist for general datasets, consistent accuracy across different domains, particularly in medical applications, remains a challenge, highlighting the need for more robust and universally applicable methods. Improved transferability scoring would significantly accelerate model selection and development across numerous fields.