Entropy Minimization without Model Collapse: Mitigating Prediction Bias in Medical Imaging
Researchers have identified a critical failure mode in test-time adaptation methods, known as model collapse, where class clusters merge and lead to prediction bias. They propose a new objective, Distribution Shift Bias Reduction (DSBR), to counteract this issue by ensuring balanced class contributions to the loss function. Experiments on medical imaging datasets and ImageNet-C demonstrate that DSBR effectively stabilizes adaptation, prevents collapse, and achieves competitive or superior performance. AI
IMPACT Mitigates a key failure mode in test-time adaptation, potentially improving reliability for AI models in critical applications like medical imaging.