机器学习期末总结见附件

我在回复（“报纸”）中，给出了若干有用的网址。可供选看。

机器学习期末体会

1. Free and good textbooks:
James, Gareth, Daniela Witten, Trevor Hastie, and Robert Tibshirani. An introduction to statistical learning. Vol. 6. New York: springer, 2013.
Hastie, Trevor, Robert Tibshirani, and Jerome Friedman. "The elements of statistical learning. 2001." NY Springer (2001).
We mainly use the first intro book. I rely on the second book for support vector machine part.
2. Suggestions about what to do
1. Read the book.
2. Watch the lecture videos (https://www.r-bloggers.com/in-de ... s-of-expert-videos/)
3. Attend a class in person is way better than doing it online especially when the professor is actively doing research in this area. Our professor is very good at illustrating the intuition through graphs. I benefit a lot from his strong sense of geometry. Online courses are typically standardized without much interaction. By contrast, Q&A in a real classroom is key to a deep understanding.
4. Do the lab work. Do not just read the lab. I enjoy using R markdown which is similar to Python notebook. I copy and paste the code from the book and then play with it a bit, and write down comments. This kind of active reading is much more productive than just reading the book.
5. Do the homework. The solutions are available online, but always try to do it by yourself first. All the theoretical questions can be solved using basic probability and calculus, no measure theory necessary. All computer questions can be solved using the techniques covered in the lab. They basically change the setup a little and encourage you to go back to the lab examples so that you reinforce what you learned in the lab.
6. Take quizzes and exams. Studying for exams and quizzes is a good way to force yourself to understand the basics. A Stanford professor seems to offer a free class through coursera. He also seems to give quizzes and exams which are helpful.
7. Do a project using real data. I like our data science professor’s rule: you have to use nontraditional data such as textual data. Doing a project basically allows you to apply the knowledge in the context of your discipline. Studying variance-bias trade-off one hundred times may not be as effective as seeing it once in your real application. For example, I get to see how poor the performance of SVM radial basis was when gamma=100 (selected by some criteria) in a real example.