Less is More: Learning Condensed Representation Promotes Both Generalization and Efficiency

2024年6月3日 10:00 - 11: 30

信管学院308会议室

Xudong Tian（田旭东） received the M.S. degree in the School of Computer Science and Technology at East China Normal University in 2021, and is currently working toward the Ph.D degree in the School of Computer Science and Technology at East China Normal University. His research interests include machine learning and computer vision, information bottleneck and multi-modal/view learning. His works are published in top-tier conferences and journals like IEEE TPAMI, CVPR, ICCV, etc.

Less is More: Learning Condensed Representation Promotes Both Generalization and Efficiency

As models grow larger and feature designs become more complex, there has been a prevailing belief in the necessity of capturing extensive information to achieve comprehensive understanding in both visual and linguistic applications. However, this idea often leads to significant redundancy, which not only increases computational costs but also compromises the models' ability to generalize. On the contrary, this talk delves into the principle of information bottleneck and demonstrates the promising potential of “reduction” in advancing human-like learning, autonomous driving, and embodied AI development.