3 ms·

ML is a great example since, as you mention your bottlenecks are likely solved by hanging your hat on Lapack and getting out. The complexity of the application

by tel 13y ago

ML is a great example since, as you mention your bottlenecks are likely solved by hanging your hat on Lapack and getting out. The complexity of the application derives from data manipulation and feature selection. The first is a definite win for higher-level languages and the second... might be, but is really too large a hodgepodge of techniques to say. I'm personally beginning to find that Haskell is really nice for this situation because it helps me separate stupid-programming-bugs from improper-algorithm-implementation-bugs very efficiently. This saves a lot of time when I'm looking at the output of a high-dimensional, non-deterministic program and wonder what might have gone wrong. And I've written some Clojure ML code that was quite close to Java speeds several years ago. For ML, you're either going to get the answer immediately or be waiting a while. As long as your constants for language choice aren't 2x-5x it's not a big deal. When everything is Fortran anyway, the constants aren't 2x-5x.

pron 13y ago
Machine learning is also, by a big margin, the most common of the areas I've mentioned and is therefore best positioned to benefit from "mass production" efficiency tools. It is moving from the master shop to the factory floor.
- tel 13y ago
  I kind of disagree with this. I think there are huge needs for "better language" kinds of improvements with ML that gives insight into how various pieces in a pipeline get combined, how tuning happens, how feature generation relates to input data. I don't consider ML to be terrifically repeatable today due to a lot of non-standard low-level tweaking that goes on at the top levels of work there. I think it's definitely moving, but I don't think it's nearly as mechanized as, say, AES is, given that the tradeoffs one needs to balance are significantly more complex. You can't just Machine Learn something as much as people may try.

by tel 13y ago

pron 13y ago

Machine learning is also, by a big margin, the most common of the areas I've mentioned and is therefore best positioned to benefit from "mass production" efficiency tools. It is moving from the master shop to the factory floor.

tel 13y ago
I kind of disagree with this. I think there are huge needs for "better language" kinds of improvements with ML that gives insight into how various pieces in a pipeline get combined, how tuning happens, how feature generation relates to input data. I don't consider ML to be terrifically repeatable today due to a lot of non-standard low-level tweaking that goes on at the top levels of work there. I think it's definitely moving, but I don't think it's nearly as mechanized as, say, AES is, given that the tradeoffs one needs to balance are significantly more complex. You can't just Machine Learn something as much as people may try.