Convexity, Logistic Loss, and the Overfitting Battle

When moving from regression to classification, reusing the exact same optimization components without evaluating their underlying geometry can break your training pipeline. Today, I realized why the M