Haoqin Tu (Isaac)    

I'm currently a junior student majoring in Information and Computing Science at Math & Physics Synthesis Class, Fuzhou University, graduating in June 2021. I'm also presently working as a student teaching fellow in Engineering Physics at Robotics and Intelligent Devices, Maynooth International Engineering College - Fuzhou University.

My research interests are but not limited to optimization, application of mathematical and physical computation and natural language processing. I’m now working in NGNLab, Tsinghua University focusing on conditional text generation and steganography under the supervison of Prof. Yongfeng Huang. Before that, I worked in a group aimed at low-rank matrix completion and sparse representation under the supervision of Prof. Shiping Wang. During this period, I learned basic algorithms of low-rank matrix completion and realised some of its applications in the recommender system. I also have broader interests in image processing, machine learning and computer vision.

I am looking for research interns in the future, where I can apply my passion and limited expertise for the greater good.

Email  /  简历 /  CV /  Github

profile photo
Selected Programs
Text Stega via Conditional Text Generation Guided by Knowledge Graph
Haoqin Tu, Zhongliang Yang, Siyu Zhang, Jinshuai Yang
supervised by Prof. Yongfeng Huang,  Tsinghua University,  Fall 2020
slide / code

Text steganography is not a tough task nowadays, but how to generate semantically/emotionally controlable long text to hide information is an unsolved problem. We modified the PHVM model which uses VAE frame in both word planning stage and word-level generation stage to make sure it can produce coherent and fluent sentences with given key tuples guided by our knowledge graph. Then we apply the Haffman encoding and Arithmetic encoding to hide the bitstream via probability output of sentence decoder in each step. For now, we achieved average 95%+ extracting precision through 10 times generations.

Intelligent Analysis of Law Cases
Fang Chen, Xiao Hu, Ziwei Shi, Haoqin Tu
supervised by Prof. Yongfeng Huang,  Tsinghua University,  Fall 2020

This project aims at building a large scale law suit case dataset with detailed accusation label in Chinese , designing algorithms to extract them accurately and integrating the code into software for easy invocation. Specifically, we designed a textCNN and a regular expression model to find different accusations in each case. Our final average precision and F-score came to 83%+ and 80%+ which are much higher then the requirements of micro F-score 75%.

Analysis and Diagnosis of Osteocyte Unit Structure via ConvNet
Haoqin Tu
supervised by Prof. Wenxi Liu,  Fuzhou University,  Fall 2019
data provided by Dr.Junning ChenUniversity of Exeter,  Spring 2019
report (stage 1| stage 2) / code

By analyzing the mainstream image segmentation algorithms, our goal is to build a convolutional neural network model to accurately distinguish the boundaries of osteocyte cell slices. Based on the study of image segmentation neural network U-net, we hope to adjust the skip connection in its iconic U structure while maintaining its accuracy. Inspired by densenet, we proposed a novel method: the dense block which is similar to the one in densenet is added into the U gap. At the same time, the cross-layer and adjacent layer feature of the network are reconstructed. Besides, data visualization and other application modules are also realized. We are now focusing on network training and validation on different data sets.

Various Methods of Matrix Completion and Application
Haoqin Tu
supervised by Prof. Shiping Wang,  Fuzhou University,  Spring 2019
code( algorithms in Librec / low-rank matrix recovery )

This matrix completion project consists of two parts: one is the Python version re-implementation of algorithms in the mature recommender system Librec and observe their performance on the mainstream datasets. Algorithms have been completed for now are mostly based on the traditional matrix completion method, including PMF(probability matrix completion), FM(factorization machine) etc. The second part focused on the recovery of low-rank matrix and its application. They are realised mainly by constraining the schatten-p norm of matrix. Meanwhile, it also covers several novel methods proposed in recent years such as factor group-sparse regularization for matrix completion. Our current work is to combine these algorithms with differentiable programming.

Warehouse Management Optimization Based on ACO Algorithm
Haoqin Tu ,Hongqiang Y, Zeyuan W
supervised by Prof. Yanmei Hong, Fuzhou University, Spring 2020
paper | code

In this competition paper, we addressed the warehouse management optimization problem by looking for a picker start from a review desk to pick up goods from designated shelves and return to a review desk again with the minimum time. We abstract this problem as a solution of the shortest Hamiltonian path. On this basis, a novel solution is proposed linked to ant colony algorithm : firstly, the ACO algorithm is used to solve the problem of optimal Hamiltonian circuit, then we carried out a 'ring splitting' action according to different situations of end point in the circuit to discuss. We regarded this answer as an approximate solution to the original problem of optimal path. This method successfully resolved our problem of finding the shortest Hamiltonian path in the competition.


thanks jon and markdana!