Xijie Huang (Owen) 黄悉偈

I am a PhD Student at Hong Kong University of Science and Technology (HKUST), where I work on efficient deep learning (algorithm-hardware co-design) and human-centric computer vision. I am a member of Vision and System Design Lab (VSDL), advised by Prof. Tim Kwang-Ting CHENG. I have conducted a research internship at Microsoft Research Asia (MSRA). Previously, I work with Prof. Zhiqiang Shen from MBZUAI, Prof.Cewu Lu from SJTU, and Prof.Mani Srivastava from UCLA.

CV  /  Google Scholar  /  Twitter  /  Zhihu (知乎)  /  Github

[2024-03] I will join Snap Research as a research intern this summer. See you in Santa Monica!

[2023-09] One paper accepted by EMNLP 2023

[2023-05] Start my internship at Microsoft Research Asia (MSRA)

[2022-05] One paper accepted by ICML 2022

[2021-01] One paper accepted by TPAMI

[2020-08] Start my PhD at Hong Kong University of Science and Technology (HKUST)

[2020-06] Graduate from Shanghai Jiao Tong University (SJTU)

[2020-03] One paper accepted by CVPR 2020

[2019-06] Start my research internship at University of California, Los Angeles (UCLA)

[2019-03] One paper accepted by CVPR 2019

My research interests lie in the general area of artificial intelligence, particularly in efficient deep learning and human-centric computer vision. More concretely, My research interests focus on quantization, algorithm-hardware co-design, human-object interation, and healthcare. Representative papers are highlighted.

A novel stochastic quantization framework to learn the optimal mixed precision quantization strategy.

A novel approach to push the boundaries of few-shot CoT learning to improve LLM math reasoning capabilities. We propose a coarse-to-fine pruner as a plug-and-play module for LLMs, which first identifies crucial CoT examples from a large batch and then further prunes unimportant tokens.

A new angle through the coreset selection to improve the training efficiency of quantization-aware training. Our method can achieve an accuracy of 68.39% of 4-bit quantized ResNet-18 on the ImageNet-1K dataset with only a 10% subset, which has an absolute gain of 4.24% compared to the previous SoTA.

An analysis of the underlying difficulty of ViT quantization in the view of variation. A multi-crop knowledge distillation-based quantization method is proposed.

A genetic LUT-Approximation algorithm namely GQA-LUT that can automatically determine the parameters with quantization awareness. The results demonstrate that GQA-LUT achieves negligible degradation on the challenging semantic segmentation task for both vanilla and linear Transformer models. Besides, proposed GQA-LUT enables the employment of INT8-based LUT-Approximation that achieves an area savings of 81.3~81.7% and a power reduction of 79.3~80.2% compared to the high-precision FP/INT 32 alternatives.

A per-channel activation quantization scheme with additional scaling factors that can be reparameterized as exponential biases of weights, incurring a negligible cost. Our method can quantize both weights and activations in the Bert model to only 4-bit and achieves an average GLUE score of 80.07, which is only 3.66 lower than the full-precision model, significantly outperforming the previous state-of-the-art method that had a gap of 11.48.

A tiny accelerator for mixed-bit sparse CNNs featuring a novel scheme of single vector-based compressed sparse filter (CSF) method and single input multiple output scratch pad (SIMO SPad) to effectively compress weight and fetch the needed input activation.

A visual analytics approach to support flexible and scalable video data programming for model steering with reduced human effort.

An automatic, vision-based system for monitoring and analyzing the physical and mental well-being of senior citizens.

A label-agnostic unified federated learning framework, named FedMix, for medical image segmentation based on mixed image labels. In FedMix, each client updates the federated model by integrating and effectively making use of all available labeled data ranging from strong pixel-level labels, weak bounding box labels, to weakest image-level class labels.

A transferable knowledge learner and can be cooperated with any HOI detection models to achieve desirable results. TIN outperforms state-of-the-art HOI detection results by a great margin, verifying its efficacy and flexibility.

A large-scale knowledge base PaStaNet, which contains 7M+ PaSta annotations. And two corresponding models are proposed: first, we design a model named Activity2Vec to extract PaSta features, which aim to be general representations for various activities.

A latent fingerprint enhancement method based on the progressive generative adversarial network (GAN).

A large-scale Human Activity Knowledge Engine (HAKE) based on the human body part states to promote the activity understanding.

A framework to detect trojan backdoors in DNNs via output explanation techniques

Thanks to Barron's website template .