Supervised learning upon small datasets, aka few-shot learning, aims at dealing with situations in which only a few samples are available for training purposes. Among the primary applications, one shall refer to those who use medical records since samples positive for some illnesses are sometimes not straightforward to obtain. Prototypical Neural Networks arose as one of the most widely employed techniques for such a purpose, which maps training samples from each class onto another feature space based on prototype samples. Such examples correspond to the average of the samples available from a given category. The main shortcoming of such an approach, although simple and effective, is not to consider the fact that samples may contribute differently to the computation of the prototypes. This proposal aims to model the task of learning the importance of each training sample when estimating the prototypes as a metaheuristic optimization problem; specifically, we are interested in evolutionary techniques. The approach to be proposed will be evaluated in two main tasks: (I) general-purpose datasets and (II) endoscopy images. Besides, the proposal also intends to hold an internship at the University of the Basque Country, Spain.
News published in Agência FAPESP Newsletter about the scholarship: