Improving representations for Incremental Class Learning

One major problem towards artificial intelligence is the development of systems able to incrementally learn concepts from a stream of data in which new concepts could be learned at any time, avoiding an entire retraining phase. Moreover, in order to be robust enough, the systems must avoid - or at least reduce - the progressive forgetting of already acquired knowledge. To face these issues, this thesis aims at providing techniques able to improve the representations for Convolutional Neural Networks. The main novelty is the introduction of the Soft Nearest Neighbor Loss in the incremental class learning scenario. It helps to better represent data of different categories and this is an extreme advantage when the model has to learn new classes over time. To evaluate the model, an incremental-usage of CIFAR-100 dataset has been exploited and the experiments pointed out the strong performances of the model compared to the actual state-of-the-art.