papers & websites & githubs

Papers:
When Saliency Meets Sentiment: Understanding How Image Content Invokes Emotion and Sentiment

1. Very very good result, code available on github
   Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields

2. Deep Feature Flow for Video Recognition
  code: https://github.com/msracver/Deep-Feature-Flow
  Fast object detection and segmentation in a video. 
  22.25fps reported on the Cityscapes andImageNet VID dataset with NVIDIA K40 GPU and Intel Core i7-4790 CPU.
  The naive per-frame method runs 4.05 fps.
  Comparsion between the naive method, which can be seen as any state of the art segmentation and detection algorithom perform on every frame,
  and the proposed method, is as follwing:
                per-frame    proposed    dataset
  detection       73.9        73.1       ImageNet VID
  segmetation     71.1        69.2       Cityscapes
  I think the small loss of accuracy can be ingored.
  When I test the code, the result is not good at all.

3. One-Shot Video Object Segmentation
   code: http://www.vision.ee.ethz.ch/~cvlsegmentation/osvos/

4. Zero-Shot Learning - The Good, the Bad and the Ugly
  Many zero-shot got very very bad result in the wild, only good at the dataset the authors used.
  The author tried almost all zero shot methods and released code at  https://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-computing/research/zero-shot-learning/zero-shot-learning-the-good-the-bad-and-the-ugly/

5. Deep Image Matting
   Code(implenmented a baidu engineer): https://github.com/Joker316701882/Deep-Image-Matting
   The code is easy to run: having tensorflow prepared, download the pretrained model, then run it.
   His comment here: http://blog.leanote.com/post/calebge/Deep-Image-Matting%E5%A4%8D%E7%8E%B0%E8%BF%87%E7%A8%8B%E6%80%BB%E7%BB%93

5. Focal Loss for Dense Object Detection
   Unoffical code: https://github.com/unsky/focal-loss
   
6. Understanding Black-box Predictions via Influence Functions
   Someone's comment: https://github.com/eolecvk/InfluenceFunctions/blob/master/review.pdf
   Another comment: http://nooverfit.com/wp/icml-2017%E8%AE%BA%E6%96%87%E7%B2%BE%E9%80%891-%E7%94%A8%E5%BD%B1%E5%93%8D%E5%87%BD%E6%95%B0%E7%90%86%E8%A7%A3%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0%E4%B8%AD%E7%9A%84%E9%BB%91%E7%9B%92%E9%A2%84/
7. Convexified Convolutional Neural Networks
   official code: https://github.com/zhangyuc/CCNN

8. On Calibration of Modern Neural Networks
   [来自知乎] 也是在deep learning session, 也是百分之百的实验。不过它报告了一个俺从来没有注意过的有趣现象：
   基于深度网络的分类器输出的class label probability 都over-confident；而这种现象在浅层网络中并不存在。这后面没准有东西可以挖一挖。

9. Learned Invariant Feature Transform
   Use CNN to generate key point descriptors.
   Code: https://github.com/cvlab-epfl/LIFT

10. Person Re-identification in the Wild

11. NetVLAD: CNN architecture for weakly supervised place recognition

12. These are learning with nosiy labels:
    Making Deep Neural Networks Robust to Label Noise: a Loss Correction Approach CVPR 2017
          See my comment.
          
    Toward robustness against label noise in training deep discriminative neural networks ICLR 2017
          In this paper, their model are not suitable for unknown T in "Making Deep Neural Networks Robust to Label Noise: a Loss 
          Correction Approach". Known T is used.
          
    Learning from noisy large-scale datasets with minimal supervision CVPR 2017
          i.  Their algorithm can be explained with one picture(fig 3). But you probably need to read section 3 to understand figure 3.
          ii. From table 2(their result), I think their result is not good, since they got only 2(on AP), and 1(on MAP) point precision 
              increasement, compared to "Fine-Tuning with clean labels", which is a very simple and naive method. They have a very very 
              complex network compared to simple "Fine-Tuning with clean labels". 
              I think, compared to "Fine-Tuning with clean labels", of course, the model can get better performance, since it utilizes 
              prior knowledge. But the result is not much better than simple, naive method.
          iii. From ii, I think they didn't find the crucial factor, or didn't find a good way, to solve this problem.
          iv. On the dataset used in paper of Multi-label fashion image classification with minimal human supervision,
              their method is even worser than "Fine-Tuning with clean labels", lmao.
         
    Learning visual n-grams from web data ICCV 2017
    
    Multi-label fashion image classification with minimal human supervision ICCV 2017
           i. Almost the same models as Learning from noisy large-scale datasets with minimal supervision,
              the difference can be seen from section 4.2 or Figure 5.
           ii. Got less than 1 point improvement on metric AP & MAP, compared to "Fine-Tuning with clean labels", see Table 2.
    
    Iterative Learning with Open-set Noisy Labels CVPR 2018 spotlight
            The key point in this paper is that they uses a algorithm, based on feature distance, to detect outlier/wrong labeled image.
            Their whole algorithm can be seen in figure 3.
    
    AMBIENTGAN: GENERATIVE MODELS FROM LOSSY MEASUREMENTS ICLR 2018 ORAL
            i. This paper is the greatest one in writing I ever read till now. 
            ii. The idea is quite simple that can be said in one sentence: 
                The generator learns the mapping f: image -> nosiy image, by applying its conjectural f' to an image and 
                sending this image to discriminator.
                
                Longer explanation
                     "rather than distinguish a real image from a generated image as  in a traditional GAN, our discriminator must 
                     distinguish a real measurement from a simulated measurement of a generated image;" 
                The word "measurement" in the above sentence means a distorted image. The distorted image may lacks some patchs, or some
                pixels are set to zero, etc... The discriminator needs to classify if an image is real measurement. Real measurement is 
                groundtruth that artifically made(so these are used for training the  discriminator), simulated measurement is result of 
                performing a distortion(f(theta)) on lossless images(so these images want to fool the discriminator).

    learning from noisy single-labeled data, ICLR 2018
    
    Unsupervised Learning by Predicting Noise, ICML 2017
       Others' notes on zhihu & google.
       https://zhuanlan.zhihu.com/p/27614220
    
    unsupervised representation learning by predicting image rotations
       Learn feature representation to predict image rotation.
    
    From Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels
       i. Other approaches(correction) include treating the true labels as a latent variable and the noisy labels as an observed variable so that
       EM-like algorithms can be used to learn true label distribution of the dataset. 
       ii. Techniques to re-weight confident samples have also been proposed
       iii. Another popular approach attempts at cleaning up noisy labels.
            a.  label cleaning network with some clean labels
            b.  prune the correct samples based on softmax outputs
       iv. loss
    The traditional CNN architecture is designed for clean dataset. For nosiy dataset, I think, we can use the CNN extract features, 
    but the other parts should be changed.
       
13. Global Optimality in Neural Network Training

14. Learning from Simulated and Unsupervised Images through Adversarial Training

15. One shot
    Matching Networks for One Shot Learning
         Comment by Andrej Karpathy(OpenAI, Li FeiFei's student)
         https://github.com/karpathy/paper-notes/blob/master/matching_networks.md
 
16. deep learning & unsupervised
      Clustering with Deep Learning: Taxonomy and New Methods, rejected by ICLR 2018, but a nice survey
            comments by ICLR reviewers: https://openreview.net/forum?id=B1CEaMbR-
            
      Learning Discrete Representations via Information Maximizing Self-Augmented Training, ICML 2017 
      
      CNN-Based Joint Clustering and Representation Learning with Feature Drift Compensation for Large-Scale Image Data
            Published in: IEEE Transactions on Multimedia ( Volume: 20, Issue: 2, Feb. 2018)
 
 17. Problems in ML/DL/CV area
     Troubling Trends in Machine Learning Scholarship
         blog: https://news.cnblogs.com/n/601590/
     Winner's Curse? On Pace, Progress, and Empirical Rigor
         blog: http://www.sohu.com/a/224432485_114877
 
 18. People Count
      CVPR 2016, Single-image crowd counting via multi-column convolutional neural network
            unofficial code: https://github.com/svishwa/crowdcount-mcnn
      
      Nice summary till April 2018: http://gjy3035.github.io/reading/Awesome-Crowd-Counting.html
 
 19. VQA
      CVPR 2018, Tips and tricks for visual question answering: Learnings from the 2017 challenge
      
      ECCV 2018, Learning Visual Question Answering by Bootstrapping Hard Attention
           hard attention, where some information is selectively ignored is used instead of 
           soft attention, where information is re-weighted and aggregated, but never filtered out.
      
      CVPR 2018, Learning Answer Embeddings for Visual Question Answering
           The proposed approach takes the semantic relationships (as characterized by the embeddings) among answers into 
           consideration, instead of viewing them as independent ordinal numbers. The learned embedded function can be used 
           to embed unseen answers (in the training dataset). 
           
           Key: These properties make the approach particularly appealing for transfer learning for open-ended Visual QA, where 
           the source dataset on which the model is learned has limited overlapping with the target dataset in the space of 
           answers.
      
      CVPR 2018, Cross-Dataset Adaptation for Visual Question Answering
           cross-dataset adaptation for visual question answering, our goal is to train a Visual QA model on a source dataset 
           but apply it to another target one.
           
           The key challenge is that the two datasets are constructed differently, resulting in the cross-dataset mismatch 
           on images, questions, or answers.
           
           We overcome this difficulty by proposing a novel domain adaptation algorithm. 
      
      CVPR 2018, Embodied Question Answering
           We present a new AI task – Embodied Question Answering (EmbodiedQA) – where an agent is spawned at a random 
           location in a 3D environment and asked a question (‘What color is the car?’). In order to answer, the agent must 
           first intelligently navigate to explore the environment, gather necessary visual information through 
           first-person(egocentric) vision, and then answer the question (‘orange’). 
           
           In this work, we develop a dataset of questions and answers in House3D environments [1], evaluation metrics, and 
           a hierarchical model trained with imitation and reinforcement learning.
      
      CVPR 2018, Differential Attention for Visual Question Answering
           In a word, they got better attention via complex network
           See figure 1 & figure 5 to briefly see what they do.
      
           However, the regions that the previous systems focus on are not correlated with the regions that humans focus on. 
           (The previous methods' attention is not like human.) The accuracy is limited due to this drawback.
           
           In this paper, we propose to solve this problem by using an exemplar based method. We obtain one or more supporting 
           and opposing exemplars to obtain a differential attention region. 
           
           This differential attention is closer to human attention than other image based attention methods. It also helps in 
           obtaining improved accuracy when answering questions. 
           
      CVPR 2018, Learning by Asking Questions
           Question in VQA is also learned(there is a question generator). See Figure 3.

      
      blog, VQA 之 Multimodal Compact Bilinear Pooling
      https://blog.csdn.net/bea_tree/article/details/72903566

20. Shadow Detection
      ECCV 2018, Bidirectional Feature Pyramid Network with Recurrent Attention Residual Modules for Shadow Detection
       Code on github, just modify the network.

21. 
      CVPR 2018, PoseTrack: A Benchmark for Human Pose Estimation and Tracking
      CVPR 2018, Detect-and-Track: Efficient Pose Estimation in Videos
      CVPR 2017, ArtTrack: Articulated Multi-person Tracking in the Wild
      pose flow: Efficient online pose tracking
      Simple Baselines for Human Pose Estimation and Tracking
      blog on multi methods: https://blog.csdn.net/white_xiaohao/article/details/79087565
      Comparing MASK RCNN with Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields, the architecture may doesn't
      matter as long as you get the correct core module/element/way.

22. Learning Dual Convolutional Neural Networks for Low-Level Vision

23. Interaction reasoning
      NIPS 2016, Interaction Networks for Learning about Objects, Relations and Physics
            Reasoning the 4th frame of toy data.
            reviews: https://media.nips.cc/nipsbooks/nipspapers/paper_files/nips29/reviews/2244.html
      NIPS 2017, Visual Interaction Networks
            Reasoning the 4th, 5th, 6th ... frame of toy data.
            reviews: https://media.nips.cc/nipsbooks/nipspapers/paper_files/nips30/reviews/2374.html
      NIPS 2017, A simple neural network module for relational reasoning
            The paper proposes a plug and play module (called Relation Networks (RNs)) specialized for relational reasoning. 
            The module is composed of Multi Layer Perceptrons and considers relations between all pairs of objects.
            reviews: https://media.nips.cc/nipsbooks/nipspapers/paper_files/nips30/reviews/2565.html            
      ICLR 2018, Graph Networks as Learnable Physics Engines for Inference and Control
            Using graph networks to learn physics.          
      arxiv June 2018, Relational Deep Reinforcement Learning
      arxiv June 2018, Flexible Neural Representation for Physics Prediction
      arxiv Sep 2018, Neural Allocentric Intuitive Physics Prediction from Real Videos
      
      ICML 2017, Neural Message Passing for Quantum Chemistry
            Using NN to predict chemistry properties.
            Talk on ICML 2017, https://vimeo.com/238221016
      
      digression:
      Learning to Decompose and Disentangle Representations for Video Prediction
      
24. ECCV 2018 to read
      The Devil of Face Recognition is in the Noise
      GANimation: Anatomically-aware Facial Animation from a Single Image
      Learning to Blend Photos
      Imagine This! Scripts to Compositions to Videos
    CVPR 2018 to read 
      Learning to Segment Every Thing
      LayoutNet: Reconstructing the 3D Room Layout from a Single RGB Image
      Taskonomy: Disentangling Task Transfer Learning
            https://zhuanlan.zhihu.com/p/38425434
      Cross-Domain Weakly-Supervised Object Detection through Progressive Domain Adaptation
      Unsupervised Discovery of Object Landmarks as Structural Representations
      What have we learned from deep representations for action recognition?
      Semi-parametric Image Synthesis
      VITON: An Image-based Virtual Try-on Network
   others:
      LINE ARTIST: A Multi-style Sketch to Painting Synthesis Scheme
      
25. Convolutional Neural Networks on Graphs
    PPT: http://helper.ipam.ucla.edu/publications/dlt2018/dlt2018_14506.pdf
    video on this PPT: https://www.youtube.com/watch?v=v3jZRkvIOIM&t=0s&list=WL&index=5
    a work by this guy on ICLR 2018 workshop, An Experimental Study of Neural Networks for Variable Graphs, and the review
         https://openreview.net/forum?id=SJexcZc8G

26. Placehoder title:) 
   ICLR 2018, i-RevNet: Deep Invertible Networks
      reviews: https://openreview.net/forum?id=HJsjkMb0Z
   ICLR 2019 under review, The Forward-Backward Embedding of Directed Graphs
      Code for this paper is available on github :)
   ICLR 2019 under review, Invariance and Inverse Stability under ReLU
      
Githubs:
Openai's platform that you can write your code to play game, then test with this platform. I think it can reduce a lot of your time.
https://github.com/openai/universe

Neural Complete
Use a network to generate another network, the generated network is in Python keras code.
https://github.com/kootenpv/neural_complete


Websites:
1. 如何学习统计学，或我的学习之路——初学者写给初学者
https://cosx.org/2008/11/how-to-learn-statistics-by-jthu/
讲的挺对我的胃口的，感觉这才是真的学习科学的方法。

2. Funny: "Tips on publishing in NIPS, ICML or any top tier conferences for ML"
https://www.reddit.com/r/MachineLearning/comments/3x3urc/tips_on_publishing_in_nips_icml_or_any_top_tier/

3. Changing gazes of people in photos
You can test it online. I tested with my own photo, amazing result
http://163.172.78.19/

4. deep-learning-most-amazing-applications
http://www.yaronhadad.com/deep-learning-most-amazing-applications/

5. Comments on many CVPR2017 papers
http://blog.csdn.net/zhangjunhit/article/category/6801399

6. a blog on MXNET: simple introduction
http://blog.csdn.net/u013713010/article/details/71635814http://blog.csdn.net/u013713010/article/details/71635814

7. comments on CVPR 2017
https://medium.com/@Synced/cvpr-2017-the-fusion-of-deep-learning-and-computer-vision-whats-next-f4d8e9efe2c

8. 图像拼接近年发展，介绍了各个主要方法的核心内容，写的很好
https://www.zhihu.com/question/34535199/answer/135169187

9. Machine Learning Top 10 Open Source Projects (v.Feb 2018)
https://medium.mybridge.co/machine-learning-top-10-open-source-projects-v-feb-2018-d1d39062bd20

10. CS231n Lecture 10 - Recurrent Neural Networks, Image Captioning, LSTM
Very clear, not hard to understand for new beginners, nice.
https://www.youtube.com/watch?v=iX5V1WpxxkY

11. Understanding LSTM Networks
http://colah.github.io/posts/2015-08-Understanding-LSTMs/
There is no detail equations, for the equations in the article, see
https://www.cnblogs.com/zhangchaoyang/articles/6684906.html

12. CVPR2017 Image Caption有关论文总结
https://zhuanlan.zhihu.com/p/30893160
My: 第三篇Skeleton Key: Image Captioning by Skeleton-Attribute Decomposition里面的黑体"ours"不是最好的"ours"，而是“with a removed”的一个测试，
证明他们的trick有效的。第三篇效果比第一篇效果要好的。可以对比第一篇的table 4和第三篇的table 2
My: 除了这四篇，CVPR2017还有如下image cpation文章:
Deep Reinforcement Learning-Based Image Captioning With Embedding Reward
Captioning Images With Diverse Objects
Self-Critical Sequence Training for Image Captioning

13. The Unreasonable Effectiveness of Recurrent Neural Networks
http://karpathy.github.io/2015/05/21/rnn-effectiveness/

14. A significant drawback of image caption
https://zhuanlan.zhihu.com/p/34188400

15. Very detailed explanation on Hough transformation
https://blog.csdn.net/sunshine_in_moon/article/details/45273909
clear, short, with python code
https://alyssaq.github.io/2014/understanding-hough-transform/

16. 2018-2019 International Conferences in Artificial Intelligence, Computer Vision and Image Processing
https://jackietseng.github.io/conference_call_for_paper/2018-2019-conferences.html

17. 为什么要学数学？——《魔鬼数学》[美]乔丹•艾伦伯格
it's funny
https://www.youtube.com/watch?v=9uqq2ottHnw

18. CVPR2018-Segmentation相关论文整理
very nice
https://blog.csdn.net/qq_16761599/article/details/80727385

19. CV 届的金鸡百花奖：盘点我心中的 CVPR 2018 创意 TOP10 
nice
https://www.leiphone.com/news/201807/jWMEVsMly1UigfHk.html

20. ICLR2018录用论文，七大趋势全解读
nice
http://baijiahao.baidu.com/s?id=1599776977066292834&wfr=spider&for=pc

21. Asking questions to images with deep learning: a visual-question-answering tutorial
a nice introduction tutorial though it doesn't contain latest contents.
https://blog.floydhub.com/asking-questions-to-images-with-deep-learning/

22. visual question answering papers & projects, it's up to latest 
https://github.com/handong1587/handong1587.github.io/blob/d83f99c0dbb313cdc5f3870d94e52d7bf69a95ed/_posts/deep_learning/2015-10-09-vqa.md

23. Introduction to Visual Question Answering: Datasets, Approaches and Evaluation
https://tryolabs.com/blog/2018/03/01/introduction-to-visual-question-answering/

24. Research Topics of Computer Vision & Graphics Group
https://www.hhi.fraunhofer.de/en/departments/vit/research-groups/computer-vision-graphics/research-topics.html

25. Hinton 新作「在线蒸馏」，提升深度学习分布式训练表现的利器
well written, motivation and how it works are well introduced 
http://baijiahao.baidu.com/s?id=1598625949059208402&wfr=spider&for=pc

26. 3D segmentation
a nice introduction
https://thegradient.pub/beyond-the-pixel-plane-sensing-and-learning-in-3d/

27. Online segmentation demo, SegNet
http://mi.eng.cam.ac.uk/projects/segnet/demo.php#demo

28. very short summary on many many 2017 ICLR papers 
quality 奈斯！
https://zhuanlan.zhihu.com/p/26641697

29. 关系推理水平超越人类：DeepMind展示全新神经网络推理预测技术
https://baijiahao.baidu.com/s?id=1569520267563057&wfr=spider&for=pc

30. 图深度学习(GraphDL)，下一个人工智能算法热点？一文了解最新GDL相关文章 
https://mp.weixin.qq.com/s?__biz=MzU2OTA0NzE2NA%3D%3D&mid=2247490830&idx=1&sn=6356b25bbf7e0983caad352b513c2994&scene=45#wechat_redirect
One Oxford student's note on graphDL paper
https://yobibyte.github.io/pages/paper-notes.html

31. 神经网络也可以有逻辑——解析视觉推理（Visual Reasoning） 
http://www.sohu.com/a/167168537_633698

32. 深度学习结合SLAM 语义slam 文章和代码
https://blog.csdn.net/qq_40213457/article/details/81407786

33. ECCV 2018论文解读及资源集锦（9月21日更新，含全部论文下载链接）
http://www.cvmart.net/community/article/detail/298
CVPR 2018 论文解读集锦（8月21日更新）
http://www.cvmart.net/community/article/detail/214
https://zhuanlan.zhihu.com/cvpr2018

34. 费曼谈巴西的物理教育
http://blog.sciencenet.cn/blog-1319915-1012178.html

35. SLAM的前世今生 终于有人说清楚了
全面，理论和实际(市场)都有
https://www.cnblogs.com/zengcv/p/5994587.html

36. Applications of deep learning on image processing
https://blog.csdn.net/c2a2o2/article/details/77701181

37. MIT在读博士心得：做好AI科研，你需要注意什么？
https://www.sohu.com/a/230352973_236505

38. Artificial Intelligence — The Revolution Hasn’t Happened Yet
https://medium.com/@mijordan3/artificial-intelligence-the-revolution-hasnt-happened-yet-5e1d5812e1e7

39. Capsule network
Hinton's talk at MIT. He also has this talk at other places.
https://www.youtube.com/watch?v=rTawFwUvnLE
Hinton said, "This is an amazingly good video. I wish I could explain capsules that well."
https://www.reddit.com/r/MachineLearning/comments/7ew7ba/d_capsule_networks_capsnets_tutorial/
Chinese translation of this video, in text format.
https://mp.weixin.qq.com/s/9BIbthQvePqeVdLWil7vgQ
Capsule related papers:
CVPR 2018, FACSCaps: Pose-Independent Facial Action Coding with Capsule
ECCV 2018, Neural Network Encapsulation
ICLR 2018 workshop, Spectral Capsule Networks

40.  Interpreting and Explaining Deep Models in Computer Vision, CVPR 2018 Tutorial
http://interpretable-ml.org/cvpr2018tutorial/readinglist.html

41. Deep Reinforcement Learning Doesn't Work Yet
https://www.alexirpan.com/2018/02/14/rl-hard.html
Chinese translation
https://blog.csdn.net/r1unw1w/article/details/79385925

42. Two nice answers on understanding disparity map
https://stackoverflow.com/questions/17607312/difference-between-disparity-map-and-disparity-image-in-stereo-matching

43. A short (3:13 mins) video on introducing epipolar plane images
https://www.youtube.com/watch?v=1F_5c_escis

44. Understanding the Basis of the Kalman Filter Via a Simple and Intuitive Derivation 
This paper can be found by goole it.

45. How to name mathematical variables in a paper?
https://math.stackexchange.com/questions/85963/variable-naming-convention-in-mathematical-modeling
https://math.stackexchange.com/questions/866885/is-there-a-list-of-typical-variable-letters-to-use-in-a-given-context

46. A minimal solution to hand motion capture from a single color camera at over 100fps. Easy to use, plug to run.
https://github.com/CalciferZh/minimal-hand
https://www.youtube.com/watch?v=OIRulRoBdL4

47. Time Derivative of Rotation Matrices: A Tutorial
Abstract — The time derivative of a rotation matrix equals the product of a skew-symmetric matrix and the rotation matrix itself. 
This article gives a brief tutorial on the well-known result.
https://arxiv.org/pdf/1609.06088.pdf

48. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead (HIGH QUALITY)
https://www.nature.com/articles/s42256-019-0048-x

49. Explainable Artificial Intelligence/ interpretable ai
Explainable Artificial Intelligence (XAI): Concepts, Taxonomies,Opportunities and Challenges toward Responsible AI
Explainable Artificial Intelligence: a Systematic Review
Opportunities and Challenges in ExplainableArtificial Intelligence (XAI): A Survey
https://github.com/jphall663/awesome-machine-learning-interpretability#free-books
Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead
Molnar, Christoph. Interpretable Machine Learning. Lulu. com, 2020

50. learning from delayed rewards
A PhD thsis. This paper creats Q-learning. The concepts are super-clearly explained.
This paper would teach people like me on how to write good papers but not hype papers like most CVPR papers do.
These old paper would also inspire people, because every detail is clearly explained. More knowledge would give us more imagination.
http://www.cs.rhul.ac.uk/~chrisw/new_thesis.pdf

51. TEN SIMPLE RULES FOR MATHEMATICAL WRITING
actually not only for MATHEMATICAL writting, but also natural language writing.
http://web.mit.edu/dimitrib/www/Ten_Rules.pdf

51. 图解机器学习的数学基础专辑（完结）总结：who, why, what
https://www.bilibili.com/video/BV1iW411T781?p=23

52. neural networks flops computation
https://zhuanlan.zhihu.com/p/456046786

53. Deconstructing the Homography Matrix   
Euclidean, Similarity, Affine, Projective
https://medium.com/@insight-in-plain-sight/deconstructing-the-homography-matrix-35989ecc0b2

54. Build opencv python gpu
https://forum.opencv.org/t/can-i-use-opencv-python-with-gpu/8947