Distributional Loss for Robust Classification

Researchers have proposed new loss functions for robust deep learning, with one study introducing a bimodal Gaussian distribution-based loss concept and another presenting an adaptive loss function called ALCL.

A novel loss concept for supervised classification tasks has been proposed, capturing class ambiguity and encouraging more robust decision boundaries^[1]. This approach, described in a paper submitted on 11 Jun 2026^[1], defines an optimization objective over all classifier outputs as a bimodal Gaussian distribution. Meanwhile, a separate study introduced ALCL, an adaptive loss function for robust deep learning under non-Gaussian noise, submitted on 14 Jun 2026^[2]. ALCL is a heavy-tailed loss formulation that adaptively learns its robustness geometry during optimization. The ALCL paper reported that it consistently outperforms MSE and optimally tuned generalized correntropy losses in reconstruction fidelity and downstream classification accuracy^[2]. Notably, ALCL achieved a median accuracy improvement of 4.75% on grayscale benchmarks and 4.51% on RGB datasets^[2]. The two papers present different submission dates, with the first paper stating 11 Jun 2026 and the second stating 14 Jun 2026.