EXACT: How to train your accuracy

Ivan Karpukhin, Stanislav Dereka, Sergey Kolesnikov

Abstract: Highlights•We show the limitations of surrogate losses in multi-class classification.•We propose a new optimization approach for classification tasks.•We propose a new gradient computation algorithm for the proposed loss function.•We show the benefits of the proposed method in a wide range of classification tasks.•We demonstrate robustness of the proposed optimization method to the uniform label noise.