2 ms·
It feels strange saying this but finding a function from a set of inputs and a set of outputs is what machine learning is all about.
by YeGoblynQueenne 10y ago
It feels strange saying this but finding a function from a set of inputs and a set of outputs is what machine learning is all about.
- catnaroek 10y agoNot all function spaces are equally easy to search. If you're looking for a continuous, possibly even monotonic, real map that approximates a given data set, fine, just perform a regression. If you want to attach one of finitely many tags to every element of a data set, fine, just use a clustering algorithm or perform a correspondence analysis. But, can you find higher-order functions using these techniques? Or can you find arbitrary computable functions on the integers using these techniques? Note that the integers don't have a very good notion of “sufficiently near”, unlike the reals.
- eru 10y ago> Note that the integers don't have a very good notion of “sufficiently near”, unlike the reals. You could use the p-adic norm. But that will probably not be the right notion for most applications.
- YeGoblynQueenne 10y agoMy bad, I misunderstood the intent behind this project (and I mixed it up in my mind with Barliman, elsewhere on HN at the time). This searches through a list of predefined (built-in?) functions and finds one that can match the given inputs and outputs. That sounds pretty cool and rather handy and although it's a form of machine learning (or function approximation in any case) it's not trying to sell itself off as anything it isn't, which was my initial concern. My bad for overreacting.