我是张天容，密西根大学计算机科学硕士在读，毕业于 密西根大学工程学院 and 上海交通大学密西根学院.
目前我主要希望研究自然语言处理相关领域，尤其是 language grounding。 该领域还包括了 situated language，探究语言的产生等方向。
教育经历
密西根大学 - 安娜堡分校 - 硕士
2020/09 至今 | 计算机科学
GPA: 4.00/4.00
重要课程:
重要课程:
- 信息论
- 应用GPU编程
- 近似算法
- 计算机视觉导论
密西根大学 - 安娜堡分校 - 学士
2018/09 ~ 2020/04 | 计算机科学
GPA: 3.83/4.00
重要课程:
重要课程:
- 编码与信息论
- 贝叶斯数据分析
- 数据库管理系统
- 算法导论
- 机器学习导论
- 计算机视觉中的深度学习
- 自然语言处理
- 实体人工智能中的场所相关语言学习
- 编译器原理
- 计算机架构
- University Honors
- Dean's List
上海交通大学 - 学士
2016/09 ~ 2020/08 | 电气与计算机工程
GPA: 3.01/4.00
重要课程:
重要课程:
- 现代物理
- 大数据处理
研究经历
EAGLe Lab, UMich
2019/09 至今 | 本科研究助理
Embodied Agent & Grounded LanguagE Laboratory (EAGLE) 是由 Joyce Chai 教授在密西根大学新成立的实验室。实验室主页还在建设中。实验室主要研究 language grounding 和实体 AI 相关的内容。目前正在寻找新奇尤其是与传统不同的研究课题。
Evaluation and Interpretation of Fidelity in Current VLN models (Undergoing)
Performances in Visual Language and Navigation (VLN) tasks are usually evaluated with success rate (SR) of reaching the target position. However, this measure diviates from the idea of instruction following because it lacks the supervision on the intermediate behaviours. Coverage weighted by Length Score (CLS) was introduced by Vihan Jain et. al in 2019 to account for this problem but it is a graph-based metrics that doesn't take into consideration the sementic level features of the environment. We propose a new metric that attends to both landmark and action sequencs induced by the pridicted path in hope of exposing more insightful interpretation of the current best=performing models.
Missing Step Inference in Procedural Text (2019)
Procedural text roughly resembles step decomposition of the execution of a task. The ability to complete the missing part of the procedure manifests the model's ability to reason about the causality between steps. We propose utilizing BERT-GPT2 auto-encoding scheme to predicted the abridged part of the text. The model can take image/video/knowledge graph information as external source to which the lexicons in the text are grounded.
2019/05 - 2019/08 | 本科短期研究实习
主要尝试将图神经网络添加进现有的 slot-tagger 中，希望得益于图中存在的统计信息，提升网络使用数据的效率，在冷启动时能够更加通过少量样本迅速地完成学习。
教学经历
批改作业
2020/09 ~ 2020/12 | 算法导论 | EECS 477 密西根大学
批改作业
2020/01 ~ 2020/04 | 数据库管理系统 | EECS 484 密西根大学
线上批改作业并给学生提供反馈，处理学生对阅卷的异议。
联系方式
I am Tianrong Zhang, Master's student in Computer Science at University of Michigan. I received my bachelor degrees from University of Michigan College of Engineering and Shanghai Jiao Tong University - University of Michigan Joint Institute.
Currently, I am in pursuit of research career in Natural Language Processing, with a focus on language grounding. This area of research also include topics such as situated language and emergence of language.
Education
University of Michigan, Ann Arbor - Master
Sep 2020 ~ Now | Bachelor of Science in Engineering | Computer Science
GPA: 4.00/4.00
Major Courses:
Major Courses:
- Information Theory
- Applied GPU Programming
- Fundations of Computer Vision
- Approximation Algorithms
University of Michigan, Ann Arbor - Bachelor
Sep 2018 ~ Apr 2020 | Bachelor of Science in Engineering | Computer Science
GPA: 3.83/4.00
Major Courses:
Major Courses:
- Coding Theory
- Bayes Data Analysis
- Database Management System
- Introduction to Algorithms
- Introduction to Machine Learning
- Deep Learning for Vision
- Natrual Language Processing
- Situated Language for Embodied AI
- Compiler Construction
- Computer Architecture
- University Honors
- Dean's List
Shanghai Jiao Tong University - Bachelor
Sep 2016 ~ Aug 2020 | Bachelor of Engineering | Electrical and Computer Engineering
GPA: 3.01/4.00
Major Courses:
Major Courses:
- Modern Physics
- Big Data Analysis
Research Experience
EAGLe Lab, UMich
Sep 2019 - Now | Undergraduate Research Assistant
Embodied Agent & Grounded LanguagE Laboratory (EAGLE) is a new founded laboratory led by Prof. Joyce Chai at University of Michigan. The lab webpage is still under construction. The lab focuses on language grounding as well as the embodiment of AIs and is now actively seeking for extraordinary and most importantly pattern-changing topics. As a founding member, I am engaged in forming research ideas and reviewing related works.
Evaluation and Interpretation of Fidelity in Current VLN models (Undergoing)
Performances in Visual Language and Navigation (VLN) tasks are usually evaluated with success rate (SR) of reaching the target position. However, this measure diviates from the idea of instruction following because it lacks the supervision on the intermediate behaviours. Coverage weighted by Length Score (CLS) was introduced by Vihan Jain et. al in 2019 to account for this problem but it is a graph-based metrics that doesn't take into consideration the sementic level features of the environment. We propose a new metric that attends to both landmark and action sequencs induced by the pridicted path in hope of exposing more insightful interpretation of the current best=performing models.
Missing Step Inference in Procedural Text (2019)
Procedural text roughly resembles step decomposition of the execution of a task. The ability to complete the missing part of the procedure manifests the model's ability to reason about the causality between steps. We propose utilizing BERT-GPT2 auto-encoding scheme to predicted the abridged part of the text. The model can take image/video/knowledge graph information as external source to which the lexicons in the text are grounded.
May 2019 - August 2019 | Undergraduate Research Intern
I worked on incorporating graph neural network into a slot tagger to make use of the statistical relation between slots and tags. This is supposed to improve data efficiency and cold-start performances. The project has been passed on succeeding membernces.
Teaching Working
Grader
Seq 2020 ~ Dec 2020 | Introduction to Algorithms | EECS 477 UMich
Grader
Jan 2020 ~ Apr 2020 | Database Management System | EECS 484 UMich
Grade weekly assignments online, provide feedbacks and response to regrade requests.
Contact
