Author: Boxiang (Box) Zhu, Google Scholar

❗Some content is only in Chinese, please translate into English by yourself.

Now work

<aside> 💡 LLM

NLP-main-fig-modify_1.png

Timeline-based QA system: We implemented the Timeline-based-List-Question-Answering problem, specifically the desire to generate structured QA, using a QA dataset containing time. QA pair is the following:

"question": "List all entities that owned Radisson Hotel Group, also known as Radisson Hospitality, Inc., from 2010 to 2020.",
"answers": [
            "Carlson (2010, 2011, 2012, 2013, 2014, 2015, 2016)",
            "HNA Group (2016, 2017, 2018)",
            "Jinjiang International (2018, 2019, 2020)"
        ]

<aside> 🤩 Structure

GitHub - wang-zhuoran/Timeline-based-List-Question-Answering

Our project contains multiple subtasks, you can click the different subtitles to check different Tasks:

KNN Few-shot

Fine-tuning

RAG

RAG with Time Marker

Evaluation Metrics

</aside>

</aside>

<aside> 💡 Edge AI

Edge AI.png

Edge device: Smartphone, Camera, New-device: VR, Even-automobile

AI: This usually refers to deep learning and especially visual application. Currently very interested in LLM ( Large Language Model ) as well.

Build a quick, efficient AI system in Edge device. Sensible and sufficient utilisation of resources includes training and inference.

Edge AI (You can click here for more info)

<aside> 🤩 Highlight work in VR EdgeAI

Liu J, Zhu B, Wang F, et al. CaV3: Cache-assisted Viewport Adaptive Volumetric Video Streaming[C]//2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR). IEEE, 2023: 173-183.1 (CCF-A)

Edge: VR device (Or any edge device that needs to transmit volumetric video)

AI:

  1. We propose a new cross-modal model for predicting the user's viewpoint, i.e., what the user is more likely to want to see.
  2. Reinforcement learning-based algorithms for edge-side cache replacement. It is designed to minimise network bandwidth consumption. </aside>

</aside>

Previous work:

<aside> 👀 Computer Vision

<aside> 💡 Image understanding (Work in Netease Media Group, recommended product center):

  1. For recommended system using image information (similarity)

  2. Technical: Contrast learning and deep hash.

    Self-supervised Learning

    List a Table is my research habbit ( Sample ) The list a review of positive sample design in contrast learning: List of design of sample pair

    Deep Hash Network

</aside>

<aside> 💡 Video Anomaly Detection

Mainly focus on Reconstructure-based and Prediction-based methods.

It is an important application for outlier analysis and anomaly detection.

Video Anomaly Detection

</aside>

<aside> 💡 AI+medical (Multi-lesion detection based on coronary angiography images)

Untitled

This project is my bachelor thesis, which was awarded with Excellent Bachelor Thesis in Beijing. ( Only 31 students in the University (about 0.8%) can be awarded in one year. )

Main work:

  1. The dataset is built from 0 to 1 ( Data were obtained from a leading cardiology hospital in China.) while solving the problems of data cleaning, sampling, and augmentation in the object detection algorithm.
  2. Algorithm improvement: Improvements in dealing with unbalanced datasets and hard sample problems due to the unbalanced dataset. </aside>

</aside>

<aside> 💰 Grant

<aside> 1️⃣ CrowdSensing (Network Economics)

Our users are encouraged to better participate in crowd-sensing and apply it in new satellite networks.

Privacy and Incentive auction mechanism

</aside>

</aside>

<aside> 💡 Service:

Reviewer: IWQoS 2023, IEEE Satellite 2022

</aside>