Small knowledge, big challenge! This article participates in the creation of "Essential Knowledge for Programmers".
issue
How can you get started quickly and how should you learn about AI?
A1: I have no experience of success, only failure
First of all, my main purpose in blogging is to output and force input. Since I write, I have to learn to write something. At least I had to know what was going on before I could describe it in my own words. I'm a lazy person, but I want to progress, so I came up with this compulsive method. It's not because I've had success and knowledge, I've got so much knowledge that I started talking and teaching.
Because I haven't succeeded, I don't have the experience of succeeding.
I've been working on machine learning for 3 years now, though, and I've been following it ever since it was named one of the "Top 10 Buzzwords of 2017."
In April 2018, I started recording my learning content on QQ Zone, but at that time it was just for my own viewing.
So far, I've gone through so many twists and turns, and I still haven't achieved anything, I haven't become famous, I haven't become a celebrity, I haven't won a Nobel Prize or anything like that.
Although I am unknown, these three years have not been wasted, I have worked day and night, but at least I know where my time is wasted, and I can tell you about these failures.
2. Kill pigs and poke their buttocks, each has its own way
I'm going to show you an introduction to artificial intelligence that is circulating on the Internet.
This diagram is filled with hundreds of learning paths for you to learn machine learning.
第一个也是最基本的是“矩阵和线性代数”。 其他包括 DataFrames、Extract、Transform、RegEx、pdfs...
My foundation is very poor, and my education is not high, but machine learning requires a high level of basic knowledge, at least a master's degree or above.
I don't have an advantage, so I'm going to go down this path, determined to keep my feet on the ground, one step at a time. Even if it is as difficult as obtaining Buddhist scriptures from the West, I will persevere. Although I don't have three years of graduate study, I will spend six years studying on my own.
So, I started learning about matrices. There is a course on matrices in the open class of NetEase Khan Academy, and I started to study each class.
After two months of studying on and off, I began to wonder what was the point of such a boring course, and where could I use it?
So, I got annoyed and started looking directly at the code of the AI project.
I first read the book "TensorFlow Practical Google Deep Learning Framework", and when I opened it, it was still a matrix:
I couldn't understand it, I was very depressed, and I reflected that my foundation was not solid, so I went back to learn the matrix.
After studying for a while, I went downstairs to discuss with the guy who sold donkey meat hot pot (an undergraduate entrepreneur majoring in mathematics), and I felt that I was invincible in the matrix world of the catering industry.
When I opened the book, I went beyond matrix calculations and began to talk about probability, especially various formulas, which I couldn't understand.
I went to learn probability again, because the code needs to be written according to formulas, and in order to lay the foundation, I studied probability distributions for another month.
I thought that if I had been learning at this rate, I might not have access to AI in another 30 years.
I'd better forget about it and move on with this project.
3. Don't think too much, just act
I started looking for projects, starting with actual combat, focusing on running code, running code to see the effect, and my interest grew. At this time, I found the second book, "21 Projects to Play with Deep Learning: A Practical Explanation Based on TensorFlow", and started to set up the environment and run the examples in the book.
Sometimes, I don't even know what code means, and that's okay, just thinking you'll be all there in 20 years. Now let's run the example and feel the magic. If you want to know more, you can go to "TensorFlow Practical Google Deep Learning Framework" to see the corresponding theoretical explanation.
In fact, with this goal, I don't have any psychological pressure, I used to get angry if I didn't understand something, but now I will do it according to the plan if I don't understand something.
However, the road is still bumpy, it is too difficult to run a project successfully, all kinds of environmental problems, configuration problems, sometimes it takes a week or even a month, but when you run it, you will have a false feeling: this code is written by you, this project is your work!
Perhaps, it is not an illusion, in some way, it is the fruit of your labor. It's like you've built a computer yourself, you don't need to care if the CPU is made by you.
You'll learn a lot in the process of solving problems. This knowledge is not systematic, not as complete as the 29-episode matrix course, but fragmented and only relevant to this project.
After running these examples, you won't know the details of implementing these features yet, but you'll have a higher level of understanding of deep learning. For example, when you look at GAN, you know it's a generative adversarial network, a network that fights itself, creates its own fakes, and then identifies the fakes on its own. You've run an example of auto-generating anime avatars and also know that in order to do that, you first need some anime avatars as a training set.
After running these examples, you will know the application areas of deep learning, including natural language processing, image processing, audio and video processing, and so on. When you see an automatic game, you will say that it is achieved through "reinforcement learning", with experience pools and rewards for making the right moves.
Why do you suddenly understand this? Because you spent more than a week running an example that was about "reinforcement learning", you remember it so deeply.
At this time, deep learning is equivalent to having a silhouette in our minds, although we don't know its face, we know that it is a person.
4. No, we still need to think about it.
At this moment, you think you are invincible.
In fact, you're neither: you're just running code written by someone else and can only implement fixed functionality.
I ask you, you successfully ran an example of auto-generating Shakespeare's operas, but if I asked you to auto-generate the books of Zeng Guo Fan, would you be able to do it?
You proudly think that all you have to do is replace the dataset because you tried to replace Shakespeare's data with Franklin's and it worked.
But once you change Zeng Guofan's data, it's all over. The reason is that English and Chinese have different word segmentation methods, but the principle is language-independent. You only need to change one place, but you don't know where to change.
When you want to make a change, you find that the sense of accomplishment of being able to run the previous example is actually an illusion and you can't control it at all.
Then let's understand its principle, and only after understanding the principle can we know where to do surgery and where to suture.
This goes back to the beginning of learning the basics. You'll see these annoying but inevitable formulas again, like this:
Although we are still facing the same problem and seem to be back to square one, there is actually a big difference. It's like asking you to take the college entrance examination now, although you don't know what to do yet, but you know what it means, and you know how to concentrate on studying and socializing in college.
At this time, you really need to calm down and study hard, don't get distracted, don't expand your research, just figure this out, because your energy is limited.
It can be a tough time. If you can't get through it, just give up. If you get through it, there will be more difficult ones in the future.
At this point, deep learning is like having organs in our brains, and we know exactly where our ears and eyes are and what shape they are.
5. You're not the only one who gets confused. Everyone has questions.
When you see these things above and finally understand what they mean, you suddenly find that you don't need to care about these things anymore after the 2.0 version of the framework has been upgraded, and you feel a little unhappy.
In version 1.0, we still need to care about the principles and design the matrix structure. But in version 2.0, this is no longer necessary.
At this point, you don't fully understand the old version 1.0, the new version 2.0 has arrived.
Do you want to learn?
6. Official initiatives are mostly trends
I started with tensors, graphs, and sessions in TensorFlow 1.x. Due to my limited knowledge, the 1.x ecosystem is as obscure as the formula.
However, after the release of TensorFlow 2.x, advanced APIs for beginners, especially the officially promoted keras, have made programming extremely simple, and only a few dozen lines of code are needed to complete the training and use of a neural network.
I struggled with it for a while, but finally chose the new version and recommended it to others.
There are several reasons for this, which is why the official is pushing this version:
7. Communication area
Sum up:
At the moment I am still a beginner, learning by reading English documents, just a porter, not competent as your teacher, I can only give you some detailed explanations within the scope of my blog content. Because as long as I write it, I must understand it thoroughly.
I'm going to open a communication area, firstly, to better convey people's doubts about my blog, but also so that people can communicate with each other.
For example, in the article "CNN Basic Recognition - I Want to Correct My Daughter's Homework", after uploading the code to github.
hwangato 正在跑步。
But the Uchiha didn't work out.
If two people can communicate with each other, one person can instruct others to run through the code to deepen the impression and grasp it more solidly; The other person can learn things and improve their skills with the help of others.
Then I have nothing to do.