Distributional Loss for Robust Classification
For practitioners training classifiers, this offers a simple plug-in loss that enhances robustness with minimal code changes.
The paper proposes a distributional loss that models classifier outputs as a bimodal Gaussian distribution, improving robustness and mitigating overfitting without extra labels. It achieves consistent gains, especially in low-data regimes.
This paper proposes a novel loss concept for supervised classification tasks. Rather than enforcing a direct mapping from each input sample to a single assigned label, we define an optimization objective over all classifier outputs as a bimodal Gaussian distribution. This softer target formulation implicitly captures class ambiguity, mitigates overfitting, and encourages the learning of more robust decision boundaries, all without requiring additional label information. Experimental results demonstrate consistent improvements in robustness, with particularly pronounced gains in low-data regimes, while requiring only minimal modifications to standard training pipelines.