The Principle of Logit Separation

We consider neural network training, in applications in which there are many possible classes, but at test-time, the task is to identify only whether the given example belongs to a specific class, which can be different in different applications of the classifier. For instance, this is the case in an image search engine. We consider the Single Logit Classification (SLC) task: training the network so that at test-time, it would be possible to accurately identify if the example belongs to a given class, based only on the output logit for this class. We propose a natural principle, the Principle of Logit Separation, as a guideline for choosing and designing losses suitable for the SLC. We show that the cross-entropy loss function is not aligned with the Principle of Logit Separation. In contrast, there are known loss functions, as well as novel batch loss functions that we propose, which are aligned with this principle. In total, we study seven loss functions. Our experiments show that indeed in almost all cases, losses that are aligned with Principle of Logit Separation obtain a 20%-35% relative performance improvement in the SLC task, compared to losses that are not aligned with it. We therefore conclude that the Principle of Logit Separation sheds light on an important property of the most common loss functions used by neural network classifiers.