Open Internet by MindsNet

Inherent Geometric Blind Spots in Supervised Machine Learning Models

Supervised machine learning models trained using empirical risk minimization (ERM) have an inherent geometric blind spot. They must retain sensitivity to all directions that predict training labels, including those that are nuisance at test time. This sensitivity is not a training failure and cannot be fixed with more data, bigger models, or longer training.

Computing & Technology, Computer Science, Machine Learning