Top-down’s fine for playing around. You’ll run into walls, but don’t give up — go bottom-up to get unstuck.

A little rhyme to understand the big picture of top-down vs bottom-up learning, particularly in the context of machine learning (ML).

I get a lot of questions about top-down vs bottom-up learning, particularly in the context of machine learning (ML), so here’s a little rhyme to help understand the big picture:

Top-down’s fine for playing around. You’ll run into walls, but don’t give up — go bottom-up to get unstuck.

Sure, check out cool ML papers you’re interested in, toy around with ML libraries, watch some videos about how the models work. Top-down playing around can definitely pique your curiosity and feel motivating!

But top-down playing around can only give you a surface-level understanding, and if you don’t have strong math skills, you’ll quickly hit a point where your math weakness prevents you from making further progress.

At that point, it’s no longer just about being curious and motivated. It’s also about being serious. And when you need to seriously learn math in depth, the only way to do that is bottom-up.

Why? Because math is a hierarchical skill domain, just like learning a sport or a musical instrument, and you can’t master an advanced skill or concept until you’ve mastered its component sub-skills and sub-concepts.