4 ms·

Honest question : once the technics will settle a bit, and libraries are created, what will be needed, apart from knowing that machine learning algorithm are ba

by bsaul 10y ago

Honest question : once the technics will settle a bit, and libraries are created, what will be needed, apart from knowing that machine learning algorithm are based on some kind of stasticial inference, with a few settings here and there ? I mean, we don't need a phd in image compression to create a service that streams videos. We just use libraries. Same for everything in computer science, it always end up packaged in some kind of reusable code or service, and only some specialists remain working in the field to work on marginal improvements. Why would ML be any different ?

Eridrus 10y ago
Once you've plugged your data into an ML system and gives you a classifier, are you done? What if the results are not good enough, do you just move on to another problem? If you don't just move on to another problem you will need people who know how to push these systems further.
- Namrog84 10y ago
  But at some point, there is a difference between turning dials of the systems vs implementing new novel ways of these systems. Obviously we always need new novel approaches, but how much of that should stay in phd/academic/research world and how much spreads out into engineering day to day activities? I am rather newbish in the whole space, but most everything I've seen is quite often just knowing enough to turn the right dials the right way. And I do wonder how long will it be before some of these dials can start turning themselves with more iterations with a simple, "yes that's what I want, or no that's what I don't want" Sure, it might take some more cycles for it to find optimal in this example.
  - Eridrus 10y ago
    It all comes down to ROI of going further; if the ROI is small, people won't go much further than twiddling some knobs. I think there's a reason you see people throwing a lot of analysis at finance. Understanding what models are doing and adding new features/data transformations is standard work for ML practitioners now. If you believe the deep learning hype, architecture engineering is the new feature engineering, and to the extent that that is true, industry will essentially resemble academia to some extent. Maybe not making anything groundbreaking, but mixing ideas together.
pesfandiar 10y ago
Unlike video streaming libraries, there's no certain way for people to verify the outcome of an ML function is correct (e.g. Would this model work on other datasets? Is the statistical inference correct?). You need at least some basic knowledge about the underlying math to avoid pitfalls.

by bsaul 10y ago

Eridrus 10y ago

Once you've plugged your data into an ML system and gives you a classifier, are you done? What if the results are not good enough, do you just move on to another problem? If you don't just move on to another problem you will need people who know how to push these systems further.

Namrog84 10y ago
But at some point, there is a difference between turning dials of the systems vs implementing new novel ways of these systems. Obviously we always need new novel approaches, but how much of that should stay in phd/academic/research world and how much spreads out into engineering day to day activities? I am rather newbish in the whole space, but most everything I've seen is quite often just knowing enough to turn the right dials the right way. And I do wonder how long will it be before some of these dials can start turning themselves with more iterations with a simple, "yes that's what I want, or no that's what I don't want" Sure, it might take some more cycles for it to find optimal in this example.
- Eridrus 10y ago
  It all comes down to ROI of going further; if the ROI is small, people won't go much further than twiddling some knobs. I think there's a reason you see people throwing a lot of analysis at finance. Understanding what models are doing and adding new features/data transformations is standard work for ML practitioners now. If you believe the deep learning hype, architecture engineering is the new feature engineering, and to the extent that that is true, industry will essentially resemble academia to some extent. Maybe not making anything groundbreaking, but mixing ideas together.

pesfandiar 10y ago

Unlike video streaming libraries, there's no certain way for people to verify the outcome of an ML function is correct (e.g. Would this model work on other datasets? Is the statistical inference correct?). You need at least some basic knowledge about the underlying math to avoid pitfalls.