I want to dive into Machine Learning, but every "AI Roadmap" I see starts with heavy multivariable calculus and linear algebra. I haven't touched a math book in five years. Can I start with the "top-down" approach—coding the models first—and then learn the math as I go, or will I just be a "black box" coder who doesn't understand why their model is failing?
3 answers
You can absolutely start with the coding. I’m a self-taught ML dev who started in early 2023. I used the "Fast.ai" approach, which is all about getting a model running on day one. It builds your confidence! You don't need to manually calculate a gradient descent to understand that your model is "searching" for the lowest error. However, once you get to the professional level, you'll want to go back and learn why things like "Overfitting" happen. You don't need to be a mathematician, but you do need to be comfortable with logic. I learned my math through YouTube tutorials specifically for ML, which was way more engaging than a dry textbook.
Heather, I love that "Fast.ai" recommendation. Did you find that not knowing the math made it harder for you to "tune" your models, or did the modern libraries like Scikit-learn make the hyperparameter tuning intuitive enough without the formulas?
Use the "Applied" approach. Build a project that solves a personal problem. The math you need will reveal itself when you hit a roadblock in your code.
Exactly, Samantha. Learning math in a vacuum is boring. Tyler, just start a project like a "Movie Recommender" and look up the math concepts as they pop up in the tutorials!
Brandon, at first, it was a bit of "trial and error." But honestly, for most mid-level projects, the documentation for these libraries is so good that you can figure out which knobs to turn without a PhD. The math really only becomes critical when you're trying to invent a new algorithm or work with extremely niche Deep Learning architectures. For 95% of us, the coding logic is the primary hurdle.