1.1.1 Working in research vs. working in production
I use research vs. production instead of academia vs. industry because even though academia is mostly concerned with research, research isn’t mostly done in academia. In fact, ML research nowadays is spearheaded by big corporations. See 1.1.2 Research for more details.
The first question you might want to figure out is whether you want to work in research or in production. They have very different job descriptions, requirements, hiring processes, and compensations.
The goal of research is to find the answers to fundamental questions and expand the body of theoretical knowledge. A research project usually involves using scientific methods to validate whether a hypothesis or a theory is true, without worrying about the practicality of the results.
The goal of production is to create or enhance a product. A product can be a good (e.g. a car), a service (e.g. ride-sharing service), a process (e.g. detecting whether a transaction is fraudulent), or a business insight (e.g. “to maximize profit we should increase our price 10%”).
A research project doesn’t need users, but a product does. For a product to be useful, it has many more requirements other than just performance, such as inference latency, interpretability (both to users and to developers), fairness (to all subgroups of users), adaptability to changing environment. The majority of a production team’s job might be to ensure these other requirements.
The given definitions above are, of course, handwavy at best. What’s research and what’s production in machine learning remain a heated topic of debate as of 20213. One reason for the ambiguity is that novel ideas with obvious usefulness tend to attract more researchers, and solving practical problems often requires coming up with novel ideas.
For more differences between machine learning in research and in production, see Stanford’s CS 329S, lecture 1: Understanding machine learning production.
As a candidate, if you’re unfamiliar with both and not sure whether you want to find roles in research or in production, the latter might be the smoother path. There are many more roles involving production than roles involving research.
3: One example is the argument whether GPT-3 is research. Many researchers were upset when Language Models are Few-Shot Learners (OpenAI, 2020) was awarded the best paper at NeurIPS because they didn’t consider it research.