Critical Response #3

Soohyun Kim

 

“Only five years ago (circa 2018), if an engineering student had not known about deep learning, that student might have been called a fool,

but now, if deep learning is the only thing an engineering student knows, that student is definitely called a fool.”

 

To undergrads or high school students who want to be an AI researcher...

 

Nowadays AI-related gossip is very prevalent like celebrity gossip; At this point, ChatGPT is more popular than most celebrities. And this AI-related gossip is circulated via provocative articles on the Internet that attract people's attention. Undergrads or high school students who love tech things are especially vulnerable to becoming addicted to this gossip and mistakenly believing they have got to know a lot about AI.

 

I think such articles can have two bad influences on undergrads or high school students who want to study AI (From now on, I will just call them 'students'.);

1) Because such articles focus mainly on the outcome and performance of AI models, students do not pay attention to what theoretical/methodological background lies behind the outcome.

2) Also, because students' interests get focused only on AI itself, they lose sight of the context and considerations in the background field where AI is being applied.

 

Even if such articles try to explain the background theory and methodology, it is often very over-simplified. So, it is easy for students to get superficial knowledge only (while they believe they know well!). Until like 7-10 years ago, there was a time when it became an “it worked” publication just by applying straightforward CNN-based model architectures to various fields, and one simple internet blog post was enough to understand how the model works, but those days were gone. Nowadays AI model architectures are getting too complicated mathematically for internet articles to explain in short. For students who want to be a researcher who studies AI model architectures or pure machine learning theory, I would say looking into linear algebra or multivariable calculus textbooks one more time is more helpful than just indulging in reading those internet gossip on AI! Because to understand brand new AI models or to make brand new AI models by themselves, they later have to reach the level they can flow with linear algebra or multivariable calculus techniques behind AI models. And it is the only beginning! After then they have to study advanced graph theory or even topology like manifold theory too. So, Good Luck!

 

I know most engineering students are more interested in AI applications to specific fields (background domain: art, music, sports, education, biomedical research, etc.) than in machine learning theory. In the case of such students, I think that there are many cases where they get easily attracted to the fact that it is research related to their hobby (I was also one of this kind when I was an undergrad freshman!), or it looks easier than becoming a machine learning theorist. This itself is not a wrong thing, but I think students (and also I) should be wary of such thoughts leading to an easygoing attitude. It shouldn't be considered an easier way; I think an AI application researcher needs a firm attitude to be a two-way player (which is not easy!) for both AI and application domains. Good AI application researchers should have the knowledge level and critical thinking skills of real professionals in the domain, not just at the level of a hobby, to be able to communicate with them and know what they really want and need. Students, therefore, need the determination to be an expert in the domain knowledge not only in AI application methods. For instance, when they read internet articles on AI, they have to keep their sight toward the background, the domain knowledge behind the AI application like: 'Okay now I understand how this AI application works. But I better also study what was the traditional method before the AI application. Is there some virtue we lost in the AI application compared to the traditional method? If so, how can we address it with AI? What would be possible complaints by professionals who will use this AI?' AI researchers should not pretend like a visitor who brings new culture to that domain, they really have to be one of the community of that domain first.