Meeting room 4 at AIP Nihombashi, but can also attend through Zoom (see the description for a link)
Self-organizing incremental neural networks for unsupervised and supervised continual learning
A fundamental characteristic of how humans learn is that we acquire knowledge incrementally. By contrast, the standard machine learning life cycle is characterized by a clear distinction between a learning phase and an application phase. Also, machine learning models are usually designed for stationary environments where the data-generating process is assumed to be stable over time. In many real-world applications, however, the data are non-stationary with concept drifts, i.e., sudden, unforeseeable changes of the data distribution. Continual learning denotes the machine learning paradigm that considers adaptive models capable of learning new tasks and adapting to non-stationary data. A key characteristic of a continual learner is the ability to learn new tasks without compromising previously acquired knowledge, that is, without catastrophic forgetting -- even when completely new tasks are to be learned. In this talk, I will present our latest work on self-organizing incremental neural networks (SOINN) for continual learning. SOINN+ is an unsupervised learning algorithm that can detect clusters of arbitrary shapes in noisy evolving data streams. GSOINN+ is a supervised learning algorithm that can mitigate catastrophic forgetting in sequential learning tasks from different domains.
Daniel Berrar received his PhD from the University of Ulster, UK, in 2004 and is currently a Lecturer in Statistics and Data Science at the School of Mathematics & Statistics at The Open University, UK. Before joining The Open University, he was a specially appointed Associate Professor at Tokyo Institute of Technology, Japan. His research interests are statistical machine learning, data science, and the overarching field of artificial intelligence. He is active in three different research directions: (1) statistical evaluation and selection of machine learning models; (2) high-dimensional inference and optimization; and (3) continual learning. Potential applications of his work can be found in various fields that use machine learning for data analysis and knowledge extraction from high-dimensional data, with a focus on the life and health sciences.
Public events of RIKEN Center for Advanced Intelligence Project (AIP)Join community