Three Types of Research

It’s been more than 2 years since graduating from my PhD program at U or R. During these years in Tencent AI Lab, I gradually recognize that research can generally be classified into the following 3 types of research:

• Prospective Research

This type of research basically prophecy the trend in the next 5 or even 10 years. The primary goal here is to create something that is really novel. Their main contribution is not advancing the SOTA performance for tasks, though they do have to show some positive empirical results (or even SOTA performance in some minor tasks). Typical work is like the neural network papers (like the LSTM paper) before 2000. A few papers of this type will give huge influence to their field, but most are not.

This type of research is mainly conducted in the universities.

• Goal-driven Research

This type of reseach mainly aims at advancing the SOTA performance of one of several general tasks, such as machine translation. Comparing with the 1st type, they value less on the novelty in the methodology side, and the emperical results are the deal-breaker. Though their contribution can be limited in general, they can “change the world” faster than the first type of reserch. Note that some work (e.g. BERT, GPT-3) in this type can still get huge influence.

This type of research is done mostly in a industrial research lab.

• Applied Research

Comparing with the 2nd type, this type of research primarily focuses on improving a business-specific task, such as hate speech classification for a certain online platform. They don’t care novelty at all, but there’s a high requirement on performance, and researchers have to do everything they can (including data labeling) to advance the performance. Usually their business schedule is very tight, leaving less time for the researchers to have a deep understanding on the task. But it can get accomplishments faster than the other two types.

This type of research is done mostly in a industrial production team.

So which one would you like?