From the op-ed page of the NYT on January 18, “Me and My Algorithm” by Seth Freeman, which begins:
Algorithms, as you probably know, are the computer programs that infer from your profile (in the case of Facebook) and the content of your e-mails (in the case of Gmail) [or your pattern of searching and buying on Amazon.com] your interests and preferences, enabling ads to be displayed to the customers most likely to be interested in specific products.
… The algorithms are programmed, I believe, to get to know us better over time, and rather than resent the invasion of privacy I have come to feel a grudging respect for, and even a growing sense of intimacy with, my own personal algorithm. You have to admire, for example, the inventive audacity of a program that would read an e-mail someone sent me about “Holocaust deniers” and think that I might be shopping for a Holistic Dentist.
Freeman goes on in this vein with other entertaining examples.
The term algorithm has traveled a long way from its use as a technical term in mathematics to the much broader use illustrated in Freeman’s piece.
OED2 says this about the mathematical use:
Math. A process, or set of rules, usually one expressed in algebraic notation, now used esp. in computing, machine translation and linguistics.
(with cites from 1938 on, though the item can probably be antedated.)
The original mathematical use was for an effective method or procedure,
one which reduces the solution of some class of problems to a series of rote steps which, if followed to the letter, and as far as may be necessary, is bound to:
always give some answer rather than ever give no answer; always give the right answer and never give a wrong answer; always be completed in a finite number of steps, rather than in an infinite number; work for all instances of problems of the class. (link)
But the term algorithm has been extended by relaxing various of these conditions, to the point where in many current uses, as above, it conveys a procedure of any sort, but in a specifically computational context.
It might be that the intermediate step involved a detechnicalization of the mathematical term, into an ordinary-language term referring to any kind of step-by-step procedure, as in recipes or systematic medical diagnosis (the OED documents the medical usage, from 1968 on) — anything that could, in principle, be depicted in a flowchart, even if some of the steps require human judgment or estimates of probabilities.
The Facebook/Gmail/Amazon use — in which the procedures are heavily probabilistic and also designed to alter their results in accordance with experience, to figuratively “learn” from your actions, so that they change over time — then looks like a retechnicalization of the term algorithm in a new context.
Another sighting yesterday, in a January 24 New Yorker piece “You’ve Got News” by Ken Auletta, about AOL:
The company has also designed a system called Seed, a hybrid of journalism and engineering … Seed is based on the idea that editors can figure out what stories to assign by mining data from search engines like Google and social networks like Facebook. If algorithms can tell you what people are talking about, and what they’re searching for, then you know what they want to read. (p. 36)
January 22, 2011 at 11:19 am |
Over on Facebook, Dennis Lewis commented on the content of the piece rather than theterm algorithm:
That is, the papers should be telling people things that they need to know, rather than abdicating this responsibility.
January 22, 2011 at 12:18 pm |
“The Facebook/Gmail/Amazon use — in which the procedures are heavily probabilistic and also designed to alter their results in accordance with experience, to figuratively “learn” from your actions, so that they change over time — then looks like a retechnicalization of the term algorithm in a new context.”
AI folks have been using (a formal mathematical sense of) algorithm in the context of machine learning for a while — long before Amazon anyway. The recipe-book-context you give does seem like a detechnicalisation, it’s less clear to me that the AI-context is a de-/re- case.
–IP
January 23, 2011 at 6:40 am |
The history of the terminology is still unclear to me, and it’s possible that simple extension of the term in some contexts and detechnicalization in others is all that’s going on. Some areas of computational linguistics resisted the extension of the term, preferring instead procedure or simply program; for example,
But in the area of natural language parsing, algorithm has long been used for parsers that in fact combine algorithmic parsing of the sort used for programming languages with heuristics for pruning the forest of parses that result and for improving performance with experience.
January 23, 2011 at 10:21 am |
Quite. I guess I was thinking of much simpler situations, like using backpropagation to train a neural net on fairly simple data, where it seems like the use of algorithm wouldn’t be terribly distant from the mathematical definition you gave in your original post.
“it’s possible that simple extension of the term in some contexts and detechnicalization in others is all that’s going on.”
Yes, that seems plausible.
Thanks for the ref — I’ll check it out.
–IP
January 24, 2011 at 3:04 pm |
The reference is pretty obscure and from a long time ago.
All four authors went on to distinction: Walker (the senior member) in computational linguistics, Friedman in computer science (specializing in computational linguistics), Hall (now famous as Barbara Hall Partee) in semantics, and me in syntax (among other things).
January 24, 2011 at 3:09 pm |
Yet another, from a letter to Harper’s for February 2011, about journalism schools (from Lawrence Pintak of the Edward R. Murrow College of Communication at Washington State):