My first-author paper with Prof. Weishi Shi at UNT’s Active Machine Learning Lab. We plant a 4 by 4 magenta patch on some CIFAR-100 classes, then mask it at test time, and measure how long each continual-learning method keeps leaning on the patch after it stops helping. ERI splits that into three numbers: adaptation delay, performance deficit and shortcut reliance. We benchmarked seven methods in PyTorch. Accepted at IEEE ICBAIE 2025.
The Einstellung Rigidity Index
Measures how long a model clings to a shortcut after it stops helping.
IEEE ICBAIE 2025
