This site is to serve as my note-book and to effectively communicate with my students and collaborators. Every now and then, a blog may be of interest to other researchers or teachers. Views in this blog are my own. All rights of research results and findings on this blog are reserved. See also http://youtube.com/c/hongqin @hongqin
Saturday, September 29, 2018
graph permutation, isomorphism, 3xor hypothesis
3xor hypothesis
Yuan Zhou, hardness of robust graph isomorphism, Lassere gaps, and asymmetry of random graphs
https://www.youtube.com/watch?v=gAsQsEMgWU0
Given two graphs which are almost isomorphic, is it possible to find a bijection which preserves most of the edges between the two? This is the algorithmic task of Robust Graph Isomorphism, which is a natural approximation variation of the Graph Isomorphism problem. In this talk, we show that no polynomial-time algorithm solves this problem, conditioned on Feige's Random 3XOR Hypothesis. In addition, we show that the Lasserre/SOS SDP hierarchy, the most powerful SDP hierarchy known, fails quite spectacularly on this problem: it needs a linear number of rounds to distinguish two isomorphic graphs from two far-from-isomorphic graphs. Along the way, we venture into the theory of random graphs by showing that a random graph is robustly asymmetric whp, meaning that any permutation which is close to an automorphism is itself close to the identity permutation. Joint work with Ryan O'Donnell, John Wright, and Chenggang Wu.
Wednesday, September 26, 2018
compressed sensing
https://en.wikipedia.org/wiki/Compressed_sensing
Does compressed sensing only work well in linear system?
Does compressed sensing only work well in linear system?
Friday, September 21, 2018
Tuesday, September 18, 2018
lab meeting materials, capsule networks, lookout server
Justin will lead a discussion
on Hinton’s capsule networks, and show us how to setup Lookout cluster using Jupyter
notebook.
Justin’s instruction at
Capsule Networks which are
an improvement to CNNs and can work with a fraction of the training data
because of how it learns.
The conceptual overview: https://openreview.net/pdf?id=HJWLfGWRb
The training method: https://arxiv.org/abs/1710.09829
A less academic explanation: https://medium.com/ai³-theory-practice-business/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b
See
Another
resource:
Github collections on Capsule
networks
Friday, September 14, 2018
Github pytho repot
https://mp.weixin.qq.com/s/7o2q_ICEe9XlhACl7r6Uiw
https://github.com/ageitgey/face_recognition #library
Github 本周最有趣的 10 款「机器学习」开源项目 | PaperDaily #32
在碎片化阅读充斥眼球的时代,越来越少的人会去关注每篇论文背后的探索和思考。
在这个栏目里,你会快速 get 每篇精选论文的亮点和痛点,时刻紧跟 AI 前沿成果。
点击本文底部的「阅读原文」即刻加入社区,查看更多最新论文推荐。
#世界上最简单的人脸识别库
本项目号称世界上最简单的人脸识别库,可使用 Python 和命令行进行调用。该库使用 dlib 顶尖的深度学习人脸识别技术构建,在户外脸部检测数据库基准(Labeled Faces in the Wild benchmark)上的准确率高达 99.38%。
项目链接
https://github.com/ageitgey/face_recognition
#多语言词向量 Python 库
由 Facebook 开源的多语言词向量 Python 库,提供了基于 fastText 实现的多语言词向量和大规模高质量的双语词典,包括无监督和有监督两种。其中有监督方法使用双语词典或相同的字符串,无监督的方法不使用任何并行数据。
无监督方法具体可参考 Word Translation without Parallel Data 这篇论文。
论文链接
https://www.paperweekly.site/papers/1097
项目链接
https://github.com/facebookresearch/MUSE
FoolNLTK
#中文处理工具包
本项目特点:
• 可能不是最快的开源中文分词,但很可能是最准的开源中文分词
• 基于 BiLSTM 模型训练而成
• 包含分词,词性标注,实体识别,都有比较高的准确率
• 用户自定义词典
项目链接
https://github.com/rockyzhengwu/FoolNLTK
#最擅长玩《毁灭战士》的游戏AI
本项目来自卡耐基梅隆大学,是 2017 年 VizDoom《毁灭战士》AI 死亡竞赛冠军 Arnold 的 PyTorch 开源代码。
论文链接
https://www.paperweekly.site/papers/1440
项目链接
https://github.com/glample/Arnold
#2017 VQA Challenge 第一名
本项目是 2017 VQA Challenge 第一名团队两篇论文的 PyTorch 复现。
■ 论文 | Bottom-Up and Top-Down Attention for Image Captioning and Visual Question Answering
■ 链接 | https://www.paperweekly.site/papers/754
■ 论文 | Tips and Tricks for Visual Question Answering: Learnings from the 2017 Challenge
■ 链接 | https://www.paperweekly.site/papers/1441
报告解读
项目链接
https://github.com/hengyuan-hu/bottom-up-attention-vqa
#PyTorch 版 YOLOv2
著名物体检测库 YOLOv2 的 PyTorch 版本,本项目还可以将训练好的 model 转换为适配 Caffe 2。
项目链接
https://github.com/ruiminshen/yolo2-pytorch
#基于 CNN 的台铁订票验证码辨识
本项目利用简单的 Convolutional Neural Network 来实作辨识台铁订票网站的验证码,训练集的部分以模仿验证码样式的方式来产生、另外验证集的部分则自台铁订票网站撷取,再以手动方式标记约 1000 笔。
目前验证集对于 6 码型态的验证码的单码辨识率达到 98.84%,整体辨识成功率达到 91.13%。
项目链接
https://github.com/JasonLiTW/simple-railway-captcha-solver
#用 AlphaZero 下五子棋
这是一个将 AlphaZero 算法应用在五子棋的实现,由于五子棋相比围棋或国际象棋简单得多,所以只需几个小时就可以训练出一个不错的 AI 模型。
■ 论文 | AlphaZero: Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
■ 链接 | https://www.paperweekly.site/papers/1297
■ 论文 | AlphaGo Zero: Mastering the game of Go without human knowledge
■ 链接 | https://www.paperweekly.site/papers/942
项目链接
https://github.com/junxiaosong/AlphaZero_Gomoku
#OpenAI Gym 扩展集
这是一个 OpenAI Gym 库的扩展包,实现了包括:多任务学习、迁移学习、逆增强学习等功能。
项目链接
https://github.com/Breakend/gym-extensions
#Python 深度学习框架
Myia 是一个全新的 Python 深度学习框架,具有使用简单、自动微分和性能优化的特点。
项目链接
https://github.com/mila-udem/myia
本文由 AI 学术社区 PaperWeekly 精选推荐,社区目前已覆盖自然语言处理、计算机视觉、人工智能、机器学习、数据挖掘和信息检索等研究方向,点击「阅读原文」即刻加入社区!
我是彩蛋
解锁新功能:热门职位推荐!
PaperWeekly小程序升级啦
今日arXiv√猜你喜欢√热门职位√
找全职找实习都不是问题
解锁方式
1. 识别下方二维码打开小程序
2. 用PaperWeekly社区账号进行登陆
3. 登陆后即可解锁所有功能
职位发布
请添加小助手微信(pwbot01)进行咨询
长按识别二维码,使用小程序
*点击阅读原文即可注册
关于PaperWeekly
PaperWeekly 是一个推荐、解读、讨论、报道人工智能前沿论文成果的学术平台。如果你研究或从事 AI 领域,欢迎在公众号后台点击「交流群」,小助手将把你带入 PaperWeekly 的交流群里。
- 1🇼 🇼 🇼
FoolNLTK和结巴相比如何? - 浅笑々流年
66666
Subscribe to:
Posts (Atom)