Video from

Blowing up ！【 Stanford open class 】 The latest Stanford Li Feifei cs231n Computer vision courses 【 With Chinese subtitles 】- Artificial intelligence / Deep learning / machine learning /AI_ Bili, Bili _bilibilihttps://www.bilibili.com/video/BV1G94y1R7Yh?p=2&vd_source=92ff87f1a39ee7b9636c0813c2341e1a

As a design student, Xiaobai attempts to set foot in the field of artificial intelligence , I heard this course is very OK So I came to the class . I hope to give some help to students who are not from a major .

Copyright notice ： All the pictures in this article are video screenshots , The copyright owner of courseware content should be Stanford school or course instructor . The pictures in this blog should only be used for learning , Do not use for commercial purposes , Please do not infringe the intellectual property rights of the creator .

Here are Lecture2 image classification The notes

Nearest neighbor classifier

In this course, we will use python3 and numpy, The course is accompanied by a simple introductory course , The address is as follows ：

http://cs231n.github.io/python-numpy-tutorial/

Manhattan distance -Manhuttan distance L1

​​​​​(30 Bar message ) Mathematics in machine learning —— Distance definition （ Two ）： Manhattan distance （Manhattan Distance）_von Neumann The blog of -CSDN Blog _ Machine learning Manhattan distance ​​​​​​​https://blog.csdn.net/hy592070616/article/details/121569933?spm=1001.2014.3001.5501 This article introduces in great detail , Besides Manhattan distance, there are commonly used Euclidean and cosine distances .

In general, the value of each block of the two images is subtracted , Add it up ,456 It's the difference between the two images .

 nearest neighbor clssifier The problem is that , The calculation time of the test is too long . Convolutional neural network training time is long , But the test is very fast .

K-Nearest neighbors

 K The larger the size, the smoother the boundary , In most cases, we use greater than 1 Of K value . There is no predicted value in the white area .

Euclid distance -Euclidean distance L2

Euclid is also used more in image similarity calculation . How to choose different methods ？

Video original words ：

If your input features, if the individual entries in your vector have some important meaning for you ask, then maybe somehow L1 might be a more natural fit. But if it's just a generic vector in some space and you don't know which of the different elements, you don't know what they actually mean, then maybe L2 is slightly more natural.

If your input characteristics , If the entries in your vector are important to you , Then maybe L1 Maybe more natural . But if it's just a general vector in a space and you don't know which different elements , You don't know what they actually mean , that L2 It may be a little natural .

  because L1 More affected by coordinates , So you can see that the picture on the left is more horizontal and vertical , The one on the right doesn't care about this .（ Please correct any mistakes in your personal understanding .）

The teacher did a test demo Can be used to play ： http://vision.stanford.edu/teaching/cs231n-demos/knn/

L1 Certain ratio L2 Good yao ？

The teacher gave an example ：

The teacher thinks that , It often depends on your data itself and what you want to analyze . If you want to classify employees , Different vectors have different meanings , They are paid , Ability and other different types of content . Because the direction of the vector is meaningful , In such cases L1 It might make more sense . The best solution is , Try it all , See which one is good .

Hyperparameters

  It's what we set value, distance ,K value . This is not learned from the data , It's what we think is set .

To be continued .