Zhuoyu Wu bio photo

Do not go gentle into that good night

Email

LinkedIn

Github

Google Scholar

Zhuoyu Wu

Reagan · 吴卓宇   /   PhD Student, Monash University

Portrait of Zhuoyu Wu

Efficient intelligence, from algorithms to hardware.

I am a PhD student in the CyPhi AI Lab at the Monash University, supervised by Prof. Raphael Phan, and funded by Graduate Research Excellent Scholarship. Currently, I am also a visiting PhD student at Peking University under the guidance of Prof. Daquan Zhou.

Prior to Monash, I have worked on Efficient AI algorithm design as well as the SW/HW co-design under the guidance of Assoc. Prof. Dr. Zheng Wang and Dr. Wenqi Fang at Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences(SIAT, CAS). I had a wonderful time there, collaborating with a group of talented individuals to deliver several key projects.

I care about making research reproducible and practical to deploy. As a member of IAPR Technical Committee 22 on Reproducible Research in Pattern Recognition (TC22), I welcome exchanges with researchers who share these interests and encourage interested colleagues to explore TC22’s activities.

If you are interested in any aspect of me, I am always open to discussions and academic collaborations. Feel free to reach out to me at — zhuoyu.wu [at] monash.edu

  • Monash University Current · PhD student
  • SIAT, CAS Previous · Research collaboration
  • Peking University Current · Visiting PhD student

Research Interests

  • Efficient AI
  • HW/SW co-design
  • LLM
  • AI for Science

News and Updates

A compact research log: ⚙️ efficient AI / hardware, 🖥️ GUI agents, and 🎓 personal milestones.

Scroll for all updates · 2026–2023

2026

  • ⚙️ Oct 2026: TernScope: Model-Hardware Co-Design for Efficient Ternary Endoscopic Segmentation is now available on arXiv, featuring W2A8 ternary quantization and an 88.46% reduction in compute-array area. [Preprint] [Code]
  • ⚙️ Oct 2026: AGQ: Action-Guided Ultra-Low-Bit Quantization for Vision-Language-Action Models is now available on arXiv, achieving 92.7% LIBERO success at W3 and up to 6.10× model compression. [Preprint] [Code]
  • ⚙️ Oct 2026: Our paper, “VAULT: A Secure FPGA Inference Accelerator for RRAM-Inspired Biomedical Edge Intelligence,” has been accepted for an oral presentation at the IEEE International Conference on Microelectronics (ICM 2026). VAULT explores an RRAM-inspired approach to enhancing the security of FPGA-based biomedical image segmentation at the edge.
  • ⚙️ Aug 2026: Two papers were accepted by BMVC 2026: LiteUAV-DETR, which studies scale-aware feature routing for real-time UAV detection, and Motion-Equivariant Pseudo-Video Augmentation, which targets video-style generalization from a single static polyp dataset. [LiteUAV-DETR] [Pseudo-Video Augmentation]
  • 🏅 Aug 2026: DepthPolyp received the ICPR 2026 Reproducible Research in Pattern Recognition (RRPR) Badge, following a separate positive peer-review evaluation of its reproducibility, transparency, code, and data availability. [Badge Criteria] [Paper] [Code]
  • 🎓 Aug 2026: I was selected as a recipient of the IEEE Computer Society WEIGD Student Support Fund, as one of 502 selected applicants. [Award]

  • ⚙️ Aug 2026: Lightweight Polyp Segmentation via a Gain-Aware Prediction-Space Recursive Controller was accepted by ICONIP 2026 as a Poster paper. The work introduces a compact, host-portable controller that recursively refines backbone logits with minimal deployment overhead. [Preprint] [Code]
  • ⚙️ Aug 2026: RoiMAM was accepted by APSIPA ASC 2026. This work proposes a region-of-interest medical attention mechanism for efficient vision-language understanding, enabling compact yet effective medical visual question answering. [Preprint]
  • ⚙️ Aug 2026: RIGS-Refiner was accepted by APSIPA ASC 2026. This work introduces a lightweight prediction-space refinement framework for robust colonoscopy polyp segmentation, improving segmentation quality with minimal computational overhead. [Preprint] [Code]
  • ⚙️ Jul 2026: DepthPolyp in Practice was accepted by the Sixth Workshop on Reproducible Research in Pattern Recognition (RRPR 2026). This companion work expands the technical and reproducibility details of DepthPolyp and delivers a real-time iPhone inference and deployment framework. [Program]
  • ⚙️ Jun 2026: Edge-Segstar was accepted by Biomedical Signal Processing and Control (BSPC, Q1). This work explores bidirectional self-distillation to improve efficient medical segmentation models, aiming to better balance the trilemma among model size, accuracy, and inference speed. [Paper]
  • ⚙️ May 2026: RoiMAM was released on arXiv. This work introduces a region-of-interest medical attention model for efficient vision-language understanding, focusing on lesion-relevant regions to improve medical visual question answering with a compact model. [Preprint]
  • ⚙️ Apr 2026: DepthPolyp was accepted by ICPR 2026. A pseudo-depth guided lightweight segmentation model for real-time colonoscopy, achieving 3.57M parameters, 0.86 GMACs, and over 180 FPS on mobile devices. [Paper] [Code] [HF Model] [HF Demo]
  • ⚙️ Jan 2026: EndoCaver was accepted by ICASSP 2026. This work improves segmentation under fog, blur, and glare by connecting image restoration with the visual attention used for segmentation. [Preprint]

2025

  • ⚙️ Nov 2025: FLICKER was accepted by DATE 2026. We reveal fine-grained sparsity in 3D Gaussian Splatting and design a contribution-aware reconfigurable accelerator for real-time 3DGS. [Preprint]
  • ⚙️ Oct 2025: TMU was accepted by IEEE TVLSI. This work builds near-memory tensor manipulation capability for data-move-intensive, compute-light operators. [Paper]
  • ⚙️ Sep 2025: RT-Focuser was accepted by ICTA 2025 as an Oral paper. A lightweight real-time deblurring model for edge-side vision, exceeding 140 FPS on mobile devices. [Paper] [Code]
  • 🎓 Aug 2025: I began my PhD journey at Monash University.
  • 🖥️ Jul 2025: GUI-Narrator was accepted by ACM MM 2025. This work studies how to detect and caption computer GUI actions for multimodal agents. [Paper]
  • ⚙️ Apr 2025: AttenPU was accepted by GLSVLSI 2025. An area-efficient attention processor with reconfigurable FP8 precision and bidirectional dataflow. [Paper]
  • ⚙️ Jan 2025: PR-KAN was accepted by ISCAS 2025 as an Oral paper. To the best of our knowledge, this is the first FPGA accelerator dedicated to Kolmogorov-Arnold Networks (KANs). [Paper]

Earlier

  • ⚙️ Sep 2024: Fourier-LSTM was accepted by ICTA 2024. We embed Fourier transforms into LSTM to improve DNA sequencing accuracy on portable devices. [Paper]
  • 🖥️ May 2024: GUI Action Narrator was publicly released. This work uses VLMs and pre-grounding to support GUI action understanding across platforms and applications. [Paper] [Code]
  • ⚙️ Apr 2024: Harmonizing U-Nets was accepted by Computers in Biology and Medicine (Q1, IF=7.7). A cascaded design with adaptive attention fusion for efficient low-quality OCT fluid segmentation. [Paper]
  • ⚙️ Apr 2023: COMPACT was presented at DATE 2023, focusing on precision-adjustable nonlinear activation acceleration. [Paper]

Academic Service

Technical Committee Member

IAPR TC22 · Reproducible Research in Pattern Recognition

Reviewing & Programme Committees

  • Journal:

    • Information Sciences (Q1) – Reviewer
    • Biomedical Signal Processing and Control (Q1) – Reviewer
    • MEDICAL & BIOLOGICAL ENGINEERING & COMPUTING (Q2) – Reviewer
    • Frontiers in Digital Health (Q1) – Reviewer
    • Frontiers in Cell and Developmental Biology (Q1) – Reviewer
    • Applied Computational Intelligence and Soft Computing – Reviewer
  • Conference:

    • ISCAS 2027 – Reviewer
    • ICASSP 2027 – Reviewer
    • BMVC 2026 – Reviewer
    • RRPR 2026 –Programme Committee
    • IJCNN 2027 – Reviewer
    • ACMMM Dataset Track 2026 – Programme Committee Member
    • ACMMM 2026 – Reviewer
    • MICCAI 2026 – Reviewer
    • WCCI 2026 – Reviewer
    • ICASSP 2026 – Reviewer
    • BMVC 2025 – Reviewer
    • IJCNN 2025 – Reviewer

Keynote Works

Selected projects in efficient models, practical deployment, and multimodal understanding.

TernScope architecture showing SplitLK, branch reparameterization and ternary large kernels

Architecture figure from the project repository

Model–hardware co-design

TernScope

A ternary endoscopic segmentation model co-designed with an accelerator, connecting low-bit inference with efficient hardware.

DepthPolyp demo showing a colonoscopy frame, segmentation overlay and pseudo-depth prediction

Still from the repository segmentation demo

Medical vision & deployment

DepthPolyp

Lightweight colonoscopy segmentation with pseudo-depth guidance, including ONNX inference and an open-source iPhone Core ML demo.

RT-Focuser comparison showing blurry input alongside restored output

Still from the repository deblurring demo

Real-time edge vision

RT-Focuser

A lightweight image deblurring model for real-time edge deployment, with training, inference, and mobile deployment resources.

GUI-Narrator pipeline with sampled GUI screenshots, visual prompts and action narration

Pipeline figure from the project repository

Multimodal GUI understanding

GUI-Narrator

A GUI action captioning framework and the Act2Cap dataset, using cursor detection and keyframe extraction to interpret on-screen actions.