The first Doctoral Research Seminar this semester will be on "Online Continual Learning with prototype-based methods in neuromorphic hardware" by Elvin Hajizada, Intel/ICS.

Lifelong learning is a major challenge in robotic settings where the robot must continuously learn new tasks without forgetting previously learned tasks (i.e. continual learning). This must be achieved on-the-fly by learning from a stream of signals that may have strong drift. While modern deep neural network (DNN) approaches have shown great promise, they still suffer from catastrophic forgetting, where new learning causes previously learned information to be lost. Even though methods exist to counteract this issue, long retraining times and high-power requirements render these methods infeasible for robot learning. Moreover, the mainstream batch training paradigm is not well-suited for autonomous robots with real-time learning requirements. To address these issues, I propose an alternative approach based on prototypes. Prototypes are exemplars of learned classes that can be continually adapted to incorporate new knowledge. This framework is well-suited for online continual learning, where the knowledge is stored more locally, and model capacity grows by incorporating more reference vectors. As the core of my research, I have developed the Continually Learning Prototypes (CLP) algorithm, which is capable of online continual learning. The CLP is being implemented in a spiking neural network form on the Intel Loihi chip, to benefit from neuromorphic computing. While our main use case is robotic object learning, the CLP algorithm is applicable to many tasks, both with and without combining it with a DNN feature extractor. Overall, our approach offers a promising solution to the challenges of lifelong learning in robotics. By leveraging prototypes and an online continual learning framework, we can achieve accurate and efficient learning without the limitations of traditional batch training approaches.

April 17th, 2023, 10:30 am - 11:30 am in room 2026, Karlstr. 45.