Ask HN: Isn't machine learning just a special case of curve fitting?

11 points·noobermin··6 comments

I'll be honest, I am busy doing other physics things (involving actual curve fitting) but from the little I've read in my little free time, it sounds like ML is just multilinear curve fitting. Is this true?

View on HN

Yes. But curve fitting is qualitatively different with millions of dimensions than with just a few, and it's only recently that we've gotten good results in the 1M-dimensional case. For example, modern image captioning can successfully assign 1 out of 1000 semantic labels to images from ~1M pixels. There's no interpolation in that 1M-dimensional space that corresponds to the sort of curve fitting operations that work in a few dimensions.

I see, so it's the high dimensionality of parameters to fit rather than, say, a small number of parameters for millions of points.

My understanding is that only one part of ML - regression has similarities with curve fitting while classification/clustering does not. What is the equivalent in curve fitting where the goal is to find which points belong to one specific group (of points) rather than the other (clustering)?

Curve fitting makes sense for cardinal values (countable quantities), but usually does not for ordinal (numerical, but not countable, usually indicating order - e.g. rank, review scores) or nominal values (male vs female, zip codes) because you cannot even plot these as x-values on a graph in a meaningful way.

That's like saying curve fitting is just a special case of calculus.

Is the specialization here that deep? Can you give a concrete illustrative example?

which is really a special case of arithmetic.

Reply on news.ycombinator.com