Resumen
Deep neural networks have proven to be able to learn rich internal representations, including for features that can also be used for different purposes than those the networks are originally developed for. In this paper, we are interested in exploring such ability and, to this aim, we propose a novel approach for investigating the internal behavior of networks trained for source code processing tasks. Using a simple autoencoder trained in the reconstruction of vectors representing programs (i.e., program embeddings), we first analyze the performance of the internal neurons in classifying programs according to different labeling policies inspired by real programming issues, showing that some neurons can actually detect different program properties. We then study the dynamics of the network from an information-theoretic standpoint, namely by considering the neurons as signaling systems and by computing the corresponding entropy. Further, we define a way to distinguish neurons according to their behavior, to consider them as formally associated with different abstract concepts, and through the application of nonparametric statistical tests to pairs of neurons, we look for neurons with unique (or almost unique) associated concepts, showing that the entropy value of a neuron is related to the rareness of its concept. Finally, we discuss how the proposed approaches for ranking the neurons can be generalized to different domains and applied to more sophisticated and specialized networks so as to help the research in the growing field of explainable artificial intelligence.