Uses and Abuses of the Cross-Entropy Loss: Case Studies in Modern Deep Learning
This work tackles a common but flawed practice in deep learning, offering more principled methods for researchers and practitioners using label smoothing or actor-mimic reinforcement learning.
The paper addresses the misuse of categorical cross-entropy loss for non-categorical data on the simplex, proposing probabilistically-inspired alternatives based on the continuous-categorical distribution. Through experimentation, it identifies potential performance improvements and failure modes, emphasizing the need for proper probabilistic treatment.
Modern deep learning is primarily an experimental science, in which empirical advances occasionally come at the expense of probabilistic rigor. Here we focus on one such example; namely the use of the categorical cross-entropy loss to model data that is not strictly categorical, but rather takes values on the simplex. This practice is standard in neural network architectures with label smoothing and actor-mimic reinforcement learning, amongst others. Drawing on the recently discovered continuous-categorical distribution, we propose probabilistically-inspired alternatives to these models, providing an approach that is more principled and theoretically appealing. Through careful experimentation, including an ablation study, we identify the potential for outperformance in these models, thereby highlighting the importance of a proper probabilistic treatment, as well as illustrating some of the failure modes thereof.