Digifesto

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.

Advertisements

computational institutions

As the “AI ethics” debate metastasizes in my newsfeed and scholarly circles, I’m struck by the frustrations of technologists and ethicists who seem to be speaking past each other.

While these tensions play out along disciplinary fault-lines, for example, between technologists and science and technology studies (STS), the economic motivations are more often than not below the surface.

I believe this is to some extent a problem of the nomenclature, which is again the function of the disciplinary rifts involved.

Computer scientists work, generally speaking, on the design and analysis of computational systems. Many see their work as bounded by the demands of the portability and formalizability of technology (see Selbst et al., 2019). That’s their job.

This is endlessly unsatisfying to critics of the social impact of technology. STS scholars will insist on changing the subject to “sociotechnical systems”, a term that means something very general: the assemblage of people and artifacts that are not people. This, fairly, removes focus from the computational system and embeds it in a social environment.

A goal of this kind of work seems to be to hold computational systems, as they are deployed and used socially, accountable. It must be said that once this happens, we are no longer talking about the specialized domain of computer science per se. It is a wonder why STS scholars are so often picking fights with computer scientists, when their true beef seems to be with businesses that use and deploy technology.

The AI Now Institute has attempted to rebrand the problem by discussing “AI Systems” as, roughly, those sociotechnical systems that use AI. This is one the one hand more specific–AI is a particular kind of technology, and perhaps it has particular political consequences. But their analysis of AI systems quickly overflows into sweeping claims about “the technology industry”, and it’s clear that most of their recommendations have little to do with AI, and indeed are trying, once again, to change the subject from discussion of AI as a technology (a computer science research domain) to a broader set of social and political issues that do, in fact, have their own disciplines where they have been researched for years.

The problem, really, is not that any particular conversation is not happening, or is being excluded, or is being shut down. The problem is that the engineering focused conversation about AI-as-a-technology has grown very large and become an awkward synecdoche for the rise of major corporations like Google, Apple, Amazon, Facebook, and Netflix. As these corporations fund and motivate a lot of research, there’s a question of who is going to get pieces of the big pie of opportunity these companies represent, either in terms of research grants or impact due to regulation, education, etc.

But there are so many aspects of these corporations that are neither addressed by the terms “sociotechnical system”, which is just so broad, and “AI System”, which is as broad and rarely means what you’d think it does (that the system uses AI is incidental if not unnecessary; what matters is that it’s a company operating in a core social domain via primarily technological user interfaces). Neither of these gets at the unit of analysis that’s really of interest.

An alternative: “computational institution”. Computational, in the sense of computational cognitive science and computational social science: it denotes the essential role of theory of computation and statistics in explaining the behavior of the phenomenon being studied. “Institution”, in the sense of institutional economics: the unit is a firm, which is comprised of people, their equipment, and their economic relations, to their suppliers and customers. An economic lens would immediately bring into focus “the data heist” and the “role of machines” that Nissenbaum is concerned are being left to the side.