The perils of being unhinged: On the accuracy of classifiers minimizing a noise-robust convex loss
This work highlights a critical limitation in using noise-robust convex losses for binary classification, which is important for researchers and practitioners in machine learning seeking reliable methods.
The paper investigates the accuracy of binary classifiers trained by minimizing the unhinged loss, a noise-robust convex loss, and finds that even on simple linearly separable data, this approach can yield classifiers with accuracy no better than random guessing.
Van Rooyen et al. introduced a notion of convex loss functions being robust to random classification noise, and established that the "unhinged" loss function is robust in this sense. In this note we study the accuracy of binary classifiers obtained by minimizing the unhinged loss, and observe that even for simple linearly separable data distributions, minimizing the unhinged loss may only yield a binary classifier with accuracy no better than random guessing.