When *shouldn’t* you build a machine learning system?
Luke Stark raises an interesting question, directed at “ML practitioner”:
As an “ML practitioner” in on this discussion, I’ll have a go at it.
In short, one should not build an ML system for making a class of decisions if there is already a better system for making that decision that does not use ML.
An example of a comparable system that does not use ML would be a team of human beings with spreadsheets, or a team of people employed to judge for themselves.
There are a few reasons why a non-ML system could be superior in performance to an ML system:
- The people involved could have access to more data, in the course of their lives, in more dimensions of variation, than is accessible by the machine learning system.
- The people might have more sensitized ability to make semantic distinctions, such as in words or images, than an ML system
- The problem to be solved could be a “wicked problem” that is itself over a very high-dimensional space of options, with very irregular outcomes, such that they are not amenable to various forms of, e.g., linear approximations
- The people might be judging an aspect of their own social environment, such that the outcome’s validity is socially procedural (as in the outcome of a vote, or of an auction)
These are all fine reasons not to use an ML system. On the other hand, the term “ML” has been extended, as with “AI”, to include many hybrid human-computer systems, which has led to some confusion. So, for example. crowdsourced labels of images provide useful input data to ML systems. This hybrid system might perform semantic judgments over a large scale of data, at a high speed, at a tolerable rate of accuracy. Does this system count as an ML system? Or is it a form of computational institution that rivals other ways of solving the problem, and just so happens to have a machine learning algorithm as part of its process?
Meanwhile, the research frontier of machine learning is all about trying to solve problems that previously haven’t been solved, or solved as well, as alternative kinds of systems. This means there will always be a disconnect between machine learning research, which is trying to expand what it is possible to do with machine learning, and what machine learning research should, today, be deployed. Sometimes, research is done to develop technology that is not mature enough to deploy.
We should expect that a lot of ML research is done on things that should not ultimately be deployed! That’s because until we do the research, we may not understand the problem well enough to know the consequences of deployment. There’s a real sense in which ML research is about understanding the computational contours of a problem, whereas ML industry practice is about addressing the problems customers have with an efficient solution. Often this solution is a hybrid system in which ML only plays a small part; the use of ML here is really about a change in the institutional structure, not so much a part of what service is being delivered.
On the other hand, there have been a lot of cases–search engines and social media being important ones–where the scale of data and the use of ML for processing has allowed for a qualitatively different form of product or service. These are now the big deal companies we are constantly talking about. These are pretty clearly cases of successful ML.