Davide Pirovano — UniTo #
Should we always train neural networks on subclasses? #
In classification problems, the model has to predict class labels given the data  features. However, in several datasets the classes  are naturally  organized in hierarchical structures. While classification task is defined at a given level of the class structure, labels with a finer granularity can be associated to the data points and used during training. For example, if we are interested in separating images of vehicles from images of animals, we can  train the model directly using these two labels or, alternatively, on the multi-class problem defined by the finer labels associated to the specific type of vehicle or animal, such as car, ship, dog, cat etc.   Empirical evidence suggests that the second strategy can be advantageous for performance.    Our goal is to test the generality of this effect and understand its origin using real and synthetic datasets. We will show that training on fine-grained labels does not always boost the classification performance. The optimal training strategy significantly depends on the geometric structure of the data and its relation to the labels, rather than solely relying on the granularity of the labels. Factors such as the complexity of the task, dataset size, and model capacity also significantly influence the potential advantages of training with fine-grained labels.