FD202D11-D01C-4F61-B4DC-073EBCCAAE72.jpeg

Li Bo

CS Phd student@NTU, Singapore​

libo0013[at]ntu[dot]edu[dot]sg

I update my latest publications in Google Scholar page

I update my education/career on LinkedIn.

  • Google Scholar
  • LinkedIn
  • Twitter

As a PhD student, I am lucky to have the brilliant and kind researcher Prof. Ziwei Liu as my supervisor. With his guidance, I am passionate about physical world model and autonomous machine learnin.

 

My final goal is to create autonomous agents that approximate human intelligence from perception to decision. The agent is expected to quickly learn from observations in an unsupervised manner and adapt to new environments.I have been studying topics towards the objective including 1) Domain Generalization 2) Generative Models 3) Reinforcement Learning. 

Previously I was fortunately doing research at/with

  • Microsoft Research Asia, Shanghai

  (2020-2021, relaxing WestBud office, chill colleagues.)

  • Berkeley AI Research, CA, USA

  (2019-2020, nice Bay Area weather, and roll on your golden bears!)

  (2020-2021, learn to write a paper with the ML taste.)

  • DiDi Visual Perception Team, Beijing

  (2018-2019, first internship and two papers there.)

 

I love discussion & collaboration to all kinds of problems and interesting projects.

 

Feel free to drop an email~

Sparse Mixture-of-Experts are Domain Generalizable Learners
ArXiv/Code
This paper is an initial step in exploring the impact of the backbone architecture in domain generalization. We proved that a network is more robust to distribution shifts if its architecture aligns well with the invariant correlation, which is verified on synthetic and real datasets. Based on our theoretical analysis, we proposed GMoE and demonstrated its superior performance on DomainBed. As for future directions, it is interesting to develop novel backbone architectures for DG based on algorithmic alignment and classic computer vision.
WX20220928-024608_2x_edited.jpg