Something I’ve noticed, is that we tend to talk about machine learning applications that are big. Maybe you want to replace a government system, or drive a car without human at the wheel. Maybe you want to make government more rational and efficient. All these uses are big. Big in the sense that they are expressions of an expansive vision for the role ML can and should play in our society.
Contrast this big ML with small ML. Small ML positions itself as assistive rather than agentive. Small ML learns to help you, the user. Small ML helps people make decisions, rather than making decisions for people. Small ML is interested in learning about you and what you care about.
An objection one might have at this point is that it’s not clear that small ML really exists. And I’d not necessarily disagree, if we’re talking about commercial applications. Even systems that purport to know you and to curate content around your interests don’t really care about you as a person - they care about how people what other people who seem similar to you have done in the past. At the same time, you could imagine a system that takes a more than cursory interest.
Concretely, one could imagine a program on your computer that learns when the work you’re doing is interruptable and under what circumstances, and then only delivers mail or displays alerts during if it determines that you’d like to see them at that point. Or you could imagine a recommender system that looks at features of books you’ve liked in the past, and tries to recommend you new books. This is distinct from a more general recommender system in that it recognizes that two people may like the same book for different reasons, and that those tastes may be stable, but different. The foregoing examples should not be construed as statements about these applications being easy to do, necessarily, but rather that they should be possible.
There are three observations I want to make about this distinction between Big and small ML.
A lot of Big ML in use today pretends to be small. I’ve mentioned recommender systems, I’d also include facial recognition as an an example of Big ML, as it imposes the categories (the people whose faces are identified) and in it’s exegesis imposes a universal categorization. These applications pretend to be small by presenting themselves as tools for the user. Despite this, they are not tools for you to use, but rather for you to delimit preferences for treatment within the ML system.
Big ML is an attempt to see as a particular form of rational organization. Much like James C. Scott argues in Seeing Like A State, Big ML is an attempt to make certain administrative problems more legible to political or economic power. I’m not the first person to make this argument, I would bet, so I won’t belabor it too much. Obviously, this insight also extends to computer systems more broadly. Defined through Scott’s lens, Big ML makes messy systems legible to power, whereas small ml makes messy systems legible to individual users.
Small ML is not a shorthand for good ML. Just because a ML system is small, does not mean that it is good, easy or present no ethical concerns. One could easily see small ml leading to echo chambers, for example. Furthermore, there are many engineering problems that a small ml system will need to solve. There is likely less data, and less labelled data. There is already a robust literature on interaction with intelligent agents which a designer will need to negotiate. In order to be good, small ML needs to confront these challenges.
In drawing any sort of dichotomy, it is rhetorically advisable to employ polarized distinctions. Indeed, one might conclude that I think these are strict categories. I don’t. I think that ML use cases may be bigger or smaller, possibly along different axes of consideration. This distinction piece is a proposal of a possible critical distinction, it may turn out that big is small or small is big. Assuming that it does hold up, however, I would hope that future designers design for small ML.