Today I want to talk about the concept of Reverse Learning.
It helped me to get into Microsoft, it helped me to get the best thesis award for my PhD, and it helped people that I’m coaching like Nikolaos Zavitsanos, Luis Valencia, and many others to get their dream job in Data Science.
Now, hopefully, you will agree with me that an interview process is like an exam to check whether you are a good contribution to their team. If the interviewers think that you have enough knowledge and practical experience to contribute to the team from day 1 with the AI solutions they are currently implementing, then they will hire you (you passed the exam).
Tell me if you have already gone through this process, you have read some Data Science books, done some courses, or even a Master’s degree. However, you still can’t land the job you want. Is that you?
You might be frustrated, you might be tired, and you are about to give up. But somehow you have the suspicion that there is a way to join Google, Microsoft, or whatever company you like.
The core idea that hopefully changes your belief system is that the reason that you are not getting that job you want is not because you are not valid for that company. The reason is that you are learning Data Science in the wrong order and you are learning the wrong things. You need to learn the AI that is used in the industry TODAY, not the AI that was used 20 years ago.
Now the best way I found to learn the AI that is used today is a process I call Reverse Learning: instead of starting from the basic algorithms and going to what is used today, reverse the path, prioritize the techniques used in the industry today, and go back to the basics.
Now let me tell you a story about an epiphany moment I had when I was studying for my PhD that made me go from wasting my time for 6 months to starting publishing papers straight away.
One Professor suggested I read several books related to robotics and control engineering. They were the traditional and most cited books about robotics and control engineering, with hundreds of pages and super small typography. After several months of deep study, I went to my advisor with the top technique they said in the book. He said: “See Miguel, this technique works in theory because it assumes that the theoretical model of the robot is the same as the real robot, but in practice, it does not work. So your technique is not going to work in the real robot”.
I left that meeting thinking that I had wasted 6 months of my life. I was tired and frustrated, and I thought I was not going to be able to publish anything.
Then I decided to change my approach. Instead of going through more books, I would start reading state-of-the-art papers. I started reading a technique called Learning from Demonstration, by a person called Sylvain Callinon. I read all his papers and in a month I had my first experiment with the robot using Learning from Demonstration. Soon after, my first paper was accepted. See what happened? I reversed the learning path. Instead of studying the history of robotics, I studied the robotic techniques that were used today.
If you think about it deeply, don’t you think it doesn’t make sense to try to get into a high-performance team not knowing the AI that they use today? How do you expect to be hired? How do you expect to contribute to the team from day 1?
Here is a mental exercise to help you understand why reverse learning makes so much sense. Let’s say you want to lead an army to war today. Then instead of studying modern warfare techniques, like how people are using drones to hit the enemy, you decided to study everything about Julius Cesar and the Roman legions. Cesar was arguably the best general of all time, right?
Wouldn't it be crazy to study the Roman legions if you want to lead an army today? Would you get a job in the Pentagon, or would you get a job in the History Department of Harvard? It is the same with AI.
Now some questions that you might be thinking:
Q: But if you don’t learn the basic techniques of AI, you will know nothing.
A: Notice that I’m not saying that you shouldn’t learn basic techniques like linear regression. Reverse learning is about reversing the learning path, starting from the techniques that are used today, and then going to the fundamentals. What I’m proposing is to make smarter use of your time.
Q: But I can’t learn the techniques that are used in the industry today if I don’t start from the basics.
A: Yes you can. It will take another whole letter to explain to you the methodology to do it, but in this post, you can find an explanation of one of the most effective algorithms that are used today: Gradient Boosting Decision Trees (GBDT). And notice that I didn’t need to explain linear regression or logistic regression to explain GBDT.
Q: But I’m skeptical about this method, it seems to be a learn Data Science in 5 minutes thing, and mastering Data Science takes a long time
A: Again, Reverse Learning is just reversing the order you learn, to understand the techniques that are most useful in the industry. It takes 10,000 hours to master a field. All I’m saying is that it is a better use of your time if during the first 1,000 hours you learn the techniques that are used in the industry today, so you can join Google, and then spend the other 9,000h learning AI inside Google.
I hope this new idea of Reverse Learning is useful to you. A lot of people have followed it and have been able to transition their careers to Data Science.
This letter was part of my private machine learning community where I share ideas and strategies on how to switch your career to Data Science. If you think this content is valuable, join the community.