ChatGPT is popular. If you want to avoid being replaced by AI, you need to have the ability of "meta-learning"

A while ago, ChatGPT became popular all over the Internet. ChatGPT is an intelligent chatbot program developed by OpenAI, an American artificial intelligence research company. It can learn human language, chat like humans, and answer various questions.

Soon after, Baidu's Wenxin Yiyan and Alibaba's Tongyi Qianwen were also launched for testing.

Depending on the intelligence of ChatGPT, many simple and repetitive jobs will be replaced, and for enterprises, using ChatGPT will be less costly.

Image source: Tuchong Creative

As a "working person", how can we avoid being replaced by AI?

One of the most important abilities we need to have is the ability to learn throughout our lives. In other words, we need to master a good learning methodology and constantly use it to absorb new knowledge and acquire new skills throughout our lives.

So how do we develop our learning ability? We can get some inspiration from a concept in machine learning: meta-learning.

Meta-learning is the basic learning logic of machines

Meta-learning is a direction that has developed in the field of artificial intelligence in recent years. Simply put, meta-learning is to let machines learn how to learn.

Whether it is single-task learning, multi-task learning or transfer learning, it is to train the machine to learn to complete one or more specified tasks. Through training, the machine has the ability to complete a task "at the moment".

Meta-learning is to train machines to have good learning ability. In this way, when a new task appears, although the machine cannot complete the task well "at the moment", as long as it is given a little time for training, it can complete the new task very well.

For example, after meta-learning, an image classifier trained on a training set without cats can tell whether a new photo contains a cat after seeing a few cat photos. Through meta-learning, a robot trained only on flat ground can quickly complete a given task on a hillside; through meta-learning, a game-playing AI can quickly learn how to play a game it has never played before.

Analogously to the training of people, if a person is trained using meta-learning, then the focus is not on letting him master certain specific skills, but on letting him master a good "learning methodology" through the process of learning these specific skills.

Learning methodology refers to the methods, thinking models and procedures a person uses when learning.

For example, there is a task now that requires a person to write a program in a programming language that he has not learned before. Different people complete this task in different ways.

Some people's approach is to find a book on the language, read it from beginning to end, and then do the exercises at the end of the book. After they are familiar with the language, they start writing the program according to the task.

The approach of others is: first get a general understanding of the syntax of the language, then find code with similar functions to the task on the Internet, start reading the code, and write and debug based on this code. If you encounter something you don’t understand, just look up the syntax of the language directly.

These are two different learning methodologies. In this example, the first method can be summarized as "learn first, then practice", while the second method can be summarized as "learn by doing". A good programmer usually chooses the second method.

We can see that there are two differences between "skills" and "learning methodology".

The first difference is the time point of focus. "Skills" focus on the "present", while "learning methodology" focuses on the "future". In a certain task, having the skills related to the task allows you to start the task immediately, while learning methodology does not allow you to start the task immediately, but requires you to study for a period of time first. However, people who have mastered the learning methodology can complete the task well with just a little learning.

The second difference is versatility. "Skills" are usually specific to a particular task, while "learning methodology" can be applied to many different tasks.

In the above example, vocational education institutions are more about developing skills than colleges and universities. After entering a company, people who have been trained by vocational education institutions can usually get familiar with the work immediately. Colleges and universities, on the other hand, do not pay much attention to developing "skills that can be used immediately", but rather to teaching a good "learning methodology". After students have mastered the "learning methodology", no matter what job they will do in the future, whether they have understood the specific work content, as long as they have studied for a short period of time, they will be able to master it easily.

How to judge whether a person has potential?

We can also measure it from two perspectives: "time point of focus" and "universality".

For example, if a company wants to recruit a potential employee during an interview, it should not only observe the person's performance in answering interview questions. If a company only recruits people based on the performance of interview questions, then the employees it recruits are usually those who have certain skills at the moment. If a company wants to recruit a potential employee, there is a very simple way, which is to give the interviewee a probation period and give him multiple tasks that he has never seen before during the probation period, let him explore on his own, and finally see his overall performance.

People who have good overall performance must have mastered a good learning methodology. With the help of learning methodology, they can complete multiple different tasks relatively well after a short period of study. They have more potential than those who only perform well in interviews.

My postdoctoral supervisor also used this method to recruit doctoral students. For a student who meets the basic requirements, he will give the student a research topic and one or two weeks to read relevant literature, do research, and finally write a research report. This is also to observe the scientific research potential of a student. So, how does a person find and cultivate a learning methodology that suits him or her? We can also be inspired by the training model of meta-learning.

The traditional machine learning training model is characterized by "few tasks and more training data". In order for a model to complete a specific task, it needs to be trained with a large amount of data related to the task.

The characteristics of the meta-learning training model are "many tasks and little training data". The reason for "many tasks" is that we want to obtain a general "learning methodology" applicable to multiple tasks through training, rather than a specific skill that can only complete a certain task; the reason for "little training data" is that we require this learning methodology to achieve better results with only a small amount of training data.

In meta-learning training, an initial "learning methodology" is usually given first, and then certain strategies are used to continuously adjust the learning methodology based on its performance on different tasks, and finally find a learning methodology that is most effective for all tasks on average.

We can see that the characteristics of meta-learning in machine learning, "many tasks and little training data", are very similar to the intensive learning mode of college students before final exams.

In college, if some students do not devote themselves to the classroom, they need to quickly study multiple exam subjects in just one or two weeks before the exam. To achieve this goal, you must master a good learning methodology. If a student can get high scores in many courses that he does not listen to seriously just by intensive study for two weeks before the exam, then he usually has a strong "learning methodology". This learning methodology is applicable to many different courses, allowing him to sort out and master the important concepts of a course in a short period of time, and connect them in series, and then master the key points and difficulties of the learning subjects. Training can help him master the course content well.

Therefore, the reason why universities offer so many courses, in addition to hoping that students can have extensive knowledge and some basic abilities, may also be because they want students to master a good and universal "learning methodology" by studying these different types of courses.

It's not so easy for machines to replace you

Of course, the field of machine learning is gradually finding a good "learning methodology" through data training, and we humans have summarized many effective "learning methodologies".

Everyone can try these methodologies, and if they work for you, you can use them immediately.

The article is produced by Science Popularization China-Starry Sky Project (Creation and Cultivation). Please indicate the source when reprinting.

Author: Liu Xuefeng, Associate Professor and Doctoral Supervisor at Beijing University of Aeronautics and Astronautics

Reviewer: Deng Qingquan, Associate Professor, School of Mathematics and Statistics, Central China Normal University