95.11.04，现快手推荐算法工程师，前Cornell计算机系研究生。主攻机器学习和推荐系统方向。创业者·加密货币爱好者·文学爱好者。

我喜欢计算机，商业，加密经济，它们逐渐成为了我思考的坐标系，我的工具。
而我喜欢文学，是它逐渐成为了我的触觉和我的声音。

最敬佩的作家是加缪和昆德拉。会念想着一辈子的作品，是堂吉柯德。

想像路飞那样一直拼下去，当一个人有了真正想要守护的东西，他才会真正变得强大。

“One Piece到底存不存在，我不想知道。”

或许将永远困惑，但也将永远寻找。困惑是我的真实，寻找是我的勇敢。

个人经历

2019

《激进市场》中译本审阅 (2019.9)

  • 因在豆瓣上写下英译本的书评：《哈伯格税：面向共享经济的创新？》，被中译本引进团队的编辑联系，并参与进该书中译本的审阅过程。在书序第12页的感谢名单中留下名字。
  • 了解这本书的缘由，来自于以太坊创始人Vitalik在2018年4月的一篇书评博文。Posner和Glen Weyl的新书 《Radical Markets》阐述了四项激进的实验性策略，涉及经济和政治等不同领域，分别尝试去调解新时代背景下的财产权，选民投票权，移民以及数据隐私等议题。入选2018经济学人年度推荐榜单。

基于深度学习的人眼修复 (2019.5)

  • 开发者，负责模型训练和后端搭建。Cornell Tech CSC5787 Deep Learning近满分的结课项目。
  • 算法解决的问题：基于深度学习的人眼轮廓修复。适用场景如集体照中有小伙伴闭眼需要修复，以及后期人像调整时对眼部轮廓和颜色的微调。
  • 关键技术：两阶段GAN合同训练。一个负责边缘连接，另一个负责上色。



Firis (2019.1)

  • 联合创始人，兼开发者，负责模型训练和移动端开发。
  • Firis是一款视频制作应用，让用户轻松迁移电影中的特效，扮演视频角色。比如，将「复仇者联盟」中的美国队长扮演者史蒂夫罗杰斯替换为用户，让美国队长有用户的面孔和声音。
  • 关键技术：利用GAN以及Auto-encoder来完成图像转移和修复。



2018

图片反向搜索引擎 (2018.12)

  • 开发者。算法结果在大规模图片反向搜索引擎Kaggle比赛中排名第一。
  • 算法解决的问题：利用自然语言搜索相关图像。例如，如果用户输入“一个正在玩滑板的人跳跃到空中”，该算法会从大型图像数据库中将最相关的图像进行排序，并返回搜索结果给用户。
  • 关键技术：利用Partial Least Squares Regression(PLSR)，Ridge Regression， Word2Vec，对图片特征向量和词语向量进行跨模态的建模训练和回归预测。论文地址，项目Github地址。



Look (2018.9)

  • 开发者。联合Snapchat，Cornell Tech Product Studio项目。
  • Look，是一款基于实时视频推荐在线相关时装的应用。

iit (2018.5)

  • 鲸语科技公司(whalesper.com)的第二个产品iit。我主要负责深度学习模型训练和移动端开发。
  • iit: Lets food find you before you find it. 根据用户的餐饮习惯和类别，为用户推荐附近餐厅的专属菜系。
  • 加入多伦多大学工程院孵化器UofT Hatchery，2018 NEST Cohort参赛项目。
  • 技术重点：基于具体食料图像识别的深度迁移学习。




Muma - Music AI (2018.1)

  • 联合创始人，兼开发者。
  • 进驻多伦多大学计算机学院孵化器DCSIL。
  • 受邀于2018.4.30在OCE Discovery会议上展示项目。
  • [2018.4.25更新] 微软Imagine Cup加拿大赛区finalist，项目介绍。
  • [2018.5.26更新] 微软Imagine Cup加拿大赛区第二名。媒体介绍，博客记录。

2017

Generalists and Specialists (2017.11)

  • IC2S2会议接受论文第一作者。IC2S2(4th Annual International Conference on Computational Social Science)，第四届计算社会科学年度国际会议。
  • 利用自然语言大数据分析社区结构和用户兴趣分布。
  • 技术重点：LDA，LSA，word2vec。

鲸语 - 多伦多头条资讯智能聚合 (2017.3)

  • 鲸语科技(whalesper.com)创始人，兼开发者。
  • App Store多伦多搜索结果Top5。iOS应用下载量~35K。
  • 鲸语，聚合超过200个加拿大多伦多本地生活服务消息源，同城娱乐，美食推荐，热门活动，省钱攻略第一时间精选推送，内容源涵盖微信公众号，微博，知乎，今日头条等最火社交新闻媒体。利用人工智能，我们尝试改变传统资讯阅读的模式。结合你周围的人正在阅读的消息，实时浮动热门内容；利用文本数据分析，提供所有长文本以标签，极大提升诸如微博等资讯的阅读体验和效率；随时随地的照片识别和文本分析；同时增加校园聊天机器人，充分连接实时有用的校园服务，为多伦多本地学习的留学生提供最便捷的帮助。
  • [2018.2.6更新] 进驻多伦多大学工程院孵化器UofT Hatchery，参与2018 NEST Cohort。
  • 产品首发博客链接, App Store下载链接

2016

Approximate Neighbor Search & Chatbot (2016.12)

  • 在Sysomos’s Data Science Lab的演讲稿，讨论关于大规模特征向量的搜索性能，以及在聊天机器人方向上的应用。资料链接。

Autonomous Driving (2016.12)

  • 一个计算机视觉领域的项目，重心在车辆识别和视觉角度预测。在ten-fold cross validation数据集上准确率达94%。KITTI Dataset
  • Raise the idea of “visual viewpoint” in extend of object’s actual viewpoint, to allow exposing more consistent features to classifiers. The highest accuracy rate among all undergraduate and graduate students so far.
  • 技术重点：DPM， SVM(rbf kernel)， intense image preprocessing。


Video Image SmArt collaGE (2016.8)

  • VISAGE(Video Image SmArt collaGE) aims to extract the most important scenes from a clip of social video. It composes those frames into a collage for better understanding on the video content.
  • A Computer Vision project that focuses on measuring images similarity from each frames of the video. And we designed an algorithm to measure how important one frame is from that video(roughly by how long it elapsed).
  • My Idea: openCV + FFMPEG(with ImageMagick) + python(Flask) + Jquery.

React Native EveryDay (2016.7)

  • A every day challenge for myself to hone my skills in mobile development, especially with React Native. Make a small but functional prototype with different interesting ideas implemented.
  • My Idea: React Native + redux. Github Source and some lovely demos!.

Content Recommendation System (2016.7)

  • Based on TF-IDF and cosine similarity to measure the similarity between posts, users can upload any texts to receive a matching results, which can be used for “You may also like” part in any social product.
  • Backend build with python flask, with nodejs handling user input and serving information to flask.
  • Over 66 articles from UT(University of Toronto) assistant have been processed as database seed data, and it will be updated seamlessly with authenticated upload. This system has been integrated into wechat’s public channel. Github Source

SOS(Sentiment on Sentences) (2016.6)

  • Using sentiment analysis techniques to infer user-input sentences’ sentiment on the fly. Model is pre-trained by Naive Bayes and SVM classifiers, based on twitter massive tweet data. In other words, the output sentiment results can to some extends reflect how people think over that topic currently. Super interesting!
  • Built with react native, migrated from a node.js web app. Get my hands dirty with RN’s amazing mobile support.
  • Part of my react native everyday coding challenge. Github Source.

AirLoft January (2016.3)

  • A MEAN full-stack project for a peer sharing topic. We want you to share your habits and insights with the world around you!
  • A complete javascript stack project involving server-side Express and front-side Angular.
  • Nice display! Github Source, Live Demo, Video Demo.

Tweets Ocean (2016.2)

  • A hand-on Nodejs project, using Ajax to request for tweets json file and use Bootstrap\fontawesome for pretty display.
  • Involved javascript programming both client side and server side. Desire to learn more on Nodejs.
  • A elegant display with background lighthighting and informatio alert.

NLP on tweets (2016.2)

  • A hand-on natual language processing project of up to 11000 tweets on sentiment analysis.
  • Involved python programming, part-of-speech(PoS) tags, machine learning with WEKA, and IBM Watson on BlueMix.
  • Finally accuracy of 67% on test data with Naive Bayes classifier.

Donut Hero (2016.2)

  • A html5 canvas game with fancy particle and ripple effects.
  • Fully-functional game looping logic with modular javascript files for each game components.
  • Github repo

Virtual Reality World (2016.1)

  • A virtual reality website gift for my friends’ birthday!
  • Heavily use D3.js to simulate the spining bomb objects, and use Aframe.io to build the virtual scene.
  • Github repo

I Love React (2016.1)

  • A CommonJS modular project site, initialized as npm project.
  • Introduce Greensock.js animation library and accelerate animations with GPU.
  • A interesting attempt at ScrollMagic.js
  • Github repo

2015

Emocean November (2015.12)

  • A machine learning project dedicated to recognize human facial expressions and achieves up to 82% of accuracy on The Toronto Faces Dataset.
  • Ranked the 6th out of about 1600 entries among all undergraduates and graduates on Kaggle website.
  • My Idea - Neural Nets on local features + SVM(with PCA) + KNN + Ensemble Methods(mixture of experts).
  • Link on the final paper

TalkerHub March (2015.5)

  • A social media website inspired from Twitter, based on ruby on rails, hosted with Amazon S3. Users can upload their images with filters options built-in on website. Allowing users to follow other users and see their latest posts.
  • My Idea - Rails + Bootstrap + Jquery + multiple ruby gems.

Groomie (2015.4)

  • A search engine designed specifically for instagram feeds. Users can type in certain key words and the engine will scape related feeds(latest and most-liked) from Instagram API, and display in a more compact but elegant way.
  • My Idea - Rails + Bootstrap + Jquery + multiple ruby gems.