"Machine Learning"

 The name is a misnomer. And a harmful one, because it interferes with understanding the process that is really occuring.

What is really occurring is a search of a constrained program space. Let's say you want to be able to identify images of hot dogs. You begin with a plausible program for doing so, that is able to also search the space of nearby programs that might get better results on the problem. You then (in "supervised learning") provide scores that indicate how well one of these possible programs has done on solving the problem. After doing this for some time you settle upon a program that solves the problem "well enough."

This is a great technique that can produce truly impressive results on a wide class of problems, such as identifying images of hot dogs. But notice that, except for the phrase in scare quotes, there is no "learning" in the description. Calling this "learning" is importing ideological baggage that just obscures what is really going on.


  1. Correct me if I've misinterpreted your post but your core arguement is that "machine learning" isn't real learning because what machine learning AI's really do is generate programs that attempt to solve a problem whose results are then compiled against data to see how successful they are at their jobs. The program who does the task the best is then selected. Because all the AI does is write the programs whose results are judged, it technically learns nothing.

    While I agree that AI that do this aren't really learning, most AI are indeed scored based on the performance of their program and baisically keep on rewriting and editing their programs until they are able to reach a satisfactory level. I don't see how this is different from an art student trying to teach themselves how to draw human faces via trial/error and comparing their results to photographs of real faces.

  2. Well, for instance, the art student actually knows that she is drawing faces.
    A machine “learning” program that recognizes faces is only modifying some wiring according to bit patterns it is fed. It doesn’t know anything about faces at all.

  3. Yes, but isn't our knowledge of faces or all other objects also just a certain preconfigurations of neurons inside of our brains? If you accept that as true, then isn't the artist in my example also just modifying their neural wiring based on the visual patterns the artists eyes feed the artist when they're learning to draw faces? Of course, you could argue that humans have an instinctive ability to regonize human faces so my face example doesn't work but what about hunters trying to learn how find hidden raccoon dogs or an inquistive person attempting to learn something more abstract like calculus? After all, when they first begin the learning these things, they might only know that the phrases calculus and raccoon dog exist but don't know what they mean/are or why they're important. I don't see how this is different from an AI not "knowing" what a face is and slowing being able to regonize what's a face and what's not a face based on the data that its fed. If an AI is able to regonize what is and isn't a face, then how is that different from a human also being able to identify what a face is and thus knowing what a face is.

  4. Of course our knowledge of faces is not a “configuration of neurons”! Are you trying to trick me or something?


Post a Comment

Popular posts from this blog

Dumb Jock?

Misunderstanding dynamical systems