|
| Figure 1. Idea-chasing versus taking goal-driven path and picking up tools/ideas as you go |
I came across this read by John Schulman, the short essay An Opinionated Guide to ML Research, which is old but still lands hard. The distinction that stuck with me is goal-driven versus idea-driven research. Idea-driven is the mode I've always assumed how to do things and is much of how research works, i.e. you read a paper that does X, you get an idea for doing X better, and you chase that idea. Goal-driven is different. You pick a concrete capability you actually want to exist, then you solve whatever subproblems get you closer to it, borrowing methods when they help and inventing when they do not. I had never really framed basic research that way, and after sitting with it I am pretty bullish on the goal-oriented approach.
The real takeaway is that a goal keeps you focused when the resource pool for ideas is loud, and it also gives you a shot at being potentially novel ... at least I think so ... because your questions come from the capability you care about achieving rather than from the same corpus everyone else just fishes through. Schulman's point is not that ideas are worthless, just that taking such a path has the potential that you get scooped or you spend months on a 10% tweak with no larger target of value. The focus of Schulman's essay is on ML research, but it has broad applicability.
