Machine Learning, Sacred Cows, and Clausewitz
This short philippic aims at the apparent tendency of treating any tool, such as Machine Learning, as an object worth attention on its own.
Carl von Clausewitz was a famous military theorist from early 18th century Prussia. He appears in today’s discourse mostly in the form of aphorisms, in particular, “War is the continuation of Politics by other means.” That is, War should not be treated as an isolated concept, but rather, understood in the context of its triggers and outcomes, both lying in the realm of politics. I would like to paraphrase it here as “Machine Learning is the continuation of Business by other means.” (Note: this was the punchline, you should stop reading if time-constrained or disinterested already).
I firmly believe that Machine Learning (or any other hyped concept) does not solve any problem on its own. It is but a tool that despite its great usefulness must be treated in the service of the ultimate goal. In particular:
- ML does not solve self-driving cars: you actually need to manufacture the cars and their equipment, make sure sensors work in concert with the model, the legislature and ethics are taken into account, those cars bring more utility than frustration to the customers, …
- ML does not solve fake news: deciding automatedly what is or what is not fake does not help you convince people not to believe it,
- ML does not solve coronavirus: if you had a reliable model for detecting infection, you still need to deploy it at scale and put the results to practical use (which naively sounds harder than telling people “wear masks and stay at home”, at which the society has spectacularly been failing so far),
- ML does not solve churn: having a model that identifies which customers are likely to leave you does not remove the reason why they decided to leave you in the first place,
I don’t find it optimal that we actually have job positions for ML researchers or ML engineers. A more fitting and proper would be to have job positions named after the type of problems they are supposed to solve — and whether by ML or any other tool is not that important. That would, among others, require ML to be easy to apply so that years of specialised training are not required. Hopefully, recent trends such as AutoML or pre-trained models or MLaaS will keep moving us closer.
There is a gap between the actual usefulness of AI and the promise thereof, attributed to, for example:
- technical/scientific debt,
- data quality and labels,
- choosing the right model,
- mode of Operations,
- lack of a feature store
and likely more. What I, however, have seen the most is the absence of real world usecase, or its twisting and rephrasing so that it mandates a sophisticated ML use. In other words, ML for ML. As hinted in the Clausewitz paraphrasal, this is caused by ignorance of the trigger and desired outcome of ML, that is, of business.
Of course, Academia exists, seemingly independently of business, doing research to e.g. improve the tools in a general way and understand their theoretical limitations. And that is fine — I’m not saying that everything must be driven by immediate revenue, absolutely not. I’m just saying Entreprises should invest wisely and with a clearly specified desired outcome, which typically would not be basic research. It is of no surprise then that the standard process for data mining starts with business understanding, and ends with deployment. And even Academia should mind reality — c.f. here.
I don’t have any specific grudge against ML, it is just the most visible of the ignoring-business-fallacies today, but surely not the only one. For instance, it annoys me greatly to talk to Software Engineers who respond to “how would you solve this generic domain problem?” with “I don’t know, I need a technical specification, I’m just a developer”, or who don’t know why they are supposed to write a particular piece of code. There are many trends, including Big Data, Microservices, Cloudification, Functional Programming, et cetera, who deserve a similar critique, their great usefulness when used soberly notwithstanding.
Post Scriptum: unlike a lot of dudes who quote Clausewitz, I actually did read the On War book. And would not really recommend it — it is a fairly heavy read, still better than e.g. Phenomenology of Spirit or any Algebraic Topology textbook, but much worse than other classics of the genre such as those by Machiavelli or Sun-Tzu. Unless you want to be allowed to write paragraphs as this one.
