Collloquium Artificial Intelligence - Owen He, University of Groningen

Title: NEVIS: Benchmarking The Next Generation of Never-Ending Learners

Abstract:

Foundation models, such as large vision and language models, need to be constantly updated to downstream tasks and consolidate new knowledge to learn future tasks more efficiently. The continual learning, meta-learning, and auto-ML communities share the goal of designing models that can always adapt to unseen tasks and become increasingly more efficient over time by suitably transferring the accrued knowledge.Unfortunately, there was no good benchmark to investigate the question of how to efficiently adapt and consolidate knowledge.

In this talk, I will provide an overview of NEVIS, a new benchmark that consists of a stream of very challenging and diverse visual classification tasks. I will then discuss the experimental results we obtained using a variety of baseline approaches. These results found that not all our common intuitions generalize to a very large scale, and provide more insights in what approaches work better, thus shedding light on future research directions. We hope that NEVIS and our findings will help the CL, ML, and AutoML communities join forces towards the development of robust systems that can become increasingly more apt and efficient over time.

Bio:

Owen He is a Research Scientist at Google DeepMind, where he focuses on continual learning and foundation models. Prior to his work at DeepMind, Owen was a PhD student supervised by Herbert Jaeger at the University of Groningen and Jacobs University Bremen. During his PhD, Owen also visited and interned at Mila, Max Planck Institute, University of Bern, ETH Zürich, and Google DeepMind. He has previously organized the ICML2021 Continual Learning workshop and is currently serving as the publication chair of the Conference on Lifelong Learning Agents.