How Does Neural Collapse Inspire Us in Reliable Machine Learning?

[CFAR Rising Star Lecture Series]
How Does Neural Collapse Inspire Us in Reliable Machine Learning? by Dr Yang Yibo
17 Jan 2023 | 2.30pm (Singapore Time)

As deep learning achieves huge success on a wide range of application tasks, the geometric nature of the learned representation and classifier remains a mystery. A recent discovered elegant phenomenon named neural collapse describes an ideal geometric structure for the last-layer features and classifier prototypes. However, undesirable training conditions would pose challenges to such a phenomenon.

In this seminar, Dr Yang Yibo will introduce the background of neural collapse. Next, he will share how his team induces neural collapse, in both imbalanced learning and incremental learning conditions. Dr Yang will also present a novel loss function that theoretically enjoys a better convergence property than the widely adopted cross-entropy (CE) loss. Lastly, he will conclude by sharing the provable advantage and limitations of the quantum neural network over the classical neural network based on neural collapse. 

talks---yang yibo
Dr Yang Yibo
Research Scientist
JD Explore Academy

Dr Yang Yibo is currently a research scientist in JD Explore Academy. He obtained his Ph.D. degree with honours from Peking University in July 2021. His research interests lie in machine learning and computer vision. He has published more than 20 publications in top-tier journals and conferences including CVPR, NeurIPS, TPAMI, etc. Dr Yang was awarded the 2021 Wuwenjun-Outstanding Ph.D. Dissertation Award in Chinese Association for Artificial Intelligence. He has also won challenge awards including the first place in BMTT 2021, the Innovative Award in COCO 2019, and the Runner-up Award in COCO 2018.