施宇钧 (Yujun Shi)

Greetings! My name is Yujun Shi. I'm currently a PhD student in ECE, National University of Singapore, advised by Professor Vincent Y. F. Tan and Dr. Jiashi Feng. Previously, I received my B. Eng in Computer Science in Nankai University, advised by Prof. Mingming Cheng.

I'm on job market right now! Feel free to drop me an email at shi.yujun@u.nus.edu if you're interested.

Email  /  Google Scholar  /  Github  /  Twitter  /  Linkedin

Lightning fast (~1s) and accurate drag-based image editing trained on videos!

We enable "drag" editing on diffusion models. By leveraging large-scale pre-trained diffusion models, we greatly improve the generality of "drag" editing.

We find dimensional collapse of representations is one of the culprit behind the performance degradation in heterogeneous Federated Learning. We propose FedDecorr to mitigate such problem and thus facilitating FL under data heterogeneity.

We propose a Class-wise Decorrelation regularizer that enables CIL learner at initial phase to mimic representations produced by the oracle model (the model jointly trained on all classes) and thus boosting Class Incremental Learning.

Instead of only supervising the classification token in ViT, we propose a novel offline knowledge distillation method that supervises all output tokens and significantly boost ViT performance.

By injecting the missing local information into vanilla ViT with our proposed T2T module, we achieve decent ImageNet classification accuracy when training ViT from scratch.

By viewing learning algorithm as a channel (input data and output model parameters) and using chain rule of mutual information, we propose a novel algorithm called Bit-Level Information Preserving (BLIP) to combat forgetting.

By leveraging multi-level information, we achieve SOTA results on weakly-supervised instance/semantic segmentation.

template from here