Li Gu

I am a research engineer at Noah’s Ark Lab, Huawei Canada in Toronto, supervised by Prof. Wang Yang. I obtained M. Eng in Computer Engineering from University of Toronto and B. Eng in Electrical & Computer Engineering from Shanghai Jiao Tong University.

My research interests are at the intersection of computer vision and machine learning. Recently, I've been focusing on how to enable machine learning models to adapt to novel environments rapidly and efficiently, including areas in meta-learning, few-shot learning, continual learning, and prompt learning.

Email  /  Google Scholar  /  Github  /  Linkedin

Proposed a new framework for unsupervised test-time adaption toward domain shift; Formulated the adaptation process as knowledge distillation and meta-learned the scheme of knowledge aggregation from multiple source domains.

Extended ProtoNet-based few-shot image classification approaches into the video domain.

The winning team at the ORBIT few-shot object recognition challenge; Awarded cash prizes of 2,500 USD.

Proposed a bi-level optimization-based meta-learning approach to learning how to learn incrementally in the Few-Shot Class Incremental Learning(FSCIL) setting.

Proposed a replay-based Few-Shot Class Incremental Learning(FSCIL) framework that can synthesize data by a generator without storing real data during each continual learning session. Imposed entropy regularization in the generator training to encourage more uncertain examples to be utilized in knowledge distillation.

Proposed the differentiable mask-matching network (DMM-Net) for solving the video object segmentation problem where the initial object masks are provided.

Proposed an efficient framework for using a small Generative Adversarial Network (GAN) to store MCMC samples of the posterior from a large Bayesian Neural Network.