Open Internet by MindsNet

Geometric Blind Spot in Supervised Learning

Standard Empirical Risk Minimization (ERM) in machine learning leads to a geometric blind spot, causing models to be vulnerable to adversarial attacks and spurious correlations. This is a fundamental limitation of ERM, not a contingent failure. The model's internal representation is structurally forced to maintain a strictly positive Jacobian sensitivity in the direction of nuisance features.

Computing & Technology, Computer Science, Machine Learning