Gabriele Graffieti is a Ph.D. student at the University of Bologna, Italy, focusing on continual learning and deep generative models. He is also the head of AI research at AI for People, a non-profit research association that promotes the use of AI with respect for human and social needs, where he is also a part of the AI ethics team. He is one of the lead maintainers of Avalanche, the most used continual learning library and part of the PyTorch ecosystem.
In recent years, advancement in machine learning, and especially deep learning, have completely revolutionized the tech world and the society in general. Tasks that were considered impossible to be done by machines are now the foundation of many intelligent objects that surround us. However, the classical machine learning paradigm of collecting a huge amount of data, training the model, and eventually deploying it, is becoming more and more inadequate to deal with current problems. In this talk we will explore continual learning (CL), a branch of machine learning that addresses the problem of not having all the data at once during training. In this case the data is collected through time and data may contain drifts or even the task we want to solve may change over time. After an overview of the problem we analyze some modern CL algorithms and we explore the future possible research directions.