The classification of neuroimaging data, also known as brain decoding, is often limited by the lack of training data. Recently, few-shot learning methods have been developed to take advantage of deep neural networks in a context where few examples are available. Some of these methods have shown to be promising in few-shot classification of brain activation maps. They involve a pretraining phase on a generic dataset containing base classes and an adaptation phase to the few-shot problem containing novel classes. In this article, we propose several experiments to study the influence of base classes on the performance of few-shot problems.