"An exercise in inference sabotage"

March 31, 2023 @ 9:42 am · Filed by under Language and science

« previous post | next post »

Eve Armstrong, "I Murdered Conan O'Brien and Nobody Will Ever Know — an exercise in inference sabotage", 3/30/2023:

Abstract: I employ an optimization-based inference methodology together with an Ising model, in an intentionally ineffectual manner, to get away with murdering an obstreperous scientific collaborator. The antics of this collaborator, hereafter "Conan O'Brien," were impeding the publication of an important manuscript. With my tenure date looming, I found myself desperate. Luckily, I study inference, a computational means to find a solution to a physical problem, based on available measurements (say, a dead body) and a dynamical model assumed to give rise to those measurements (a murderer). If the measurements are insufficient and/or the model is incomplete, one obtains multiple "degenerate" solutions to the problem. Degenerate solutions are all equally valid given the information available, and thus render meaningless the notion of one "correct" solution. Typically in scientific research, degeneracy is undesirable. Here I describe the opposite situation: a quest to create degenerate solutions in which to cloak myself. Or even better: to render measurements incompatible with a solution in which I am the murderer. Moreover, I show how one may sabotage an inference procedure to commit an untraceable crime. I sit here now, typing victoriously, a free woman. Because you won't believe me anyway. And even if you do, you'll never prove a thing.

I have personal reasons for interest in this paper — and no, it's not because I need help planning a crime. Rather, in one of my courses this semester I'm about to take up "Hopfield networks", which are Ising models viewed as abstract models of content-addressable memory rather than abstract models of ferromagnetism. Or as Figure 2 from Eve's paper suggests,  abstract models of social dynamics:

(…and even more abstractly, similar sets of coupled non-linear differential equations can be used to model things like musical coordination, text-setting, and linguistic sound systems…)

Previously-cited works by Dr. Armstrong include

You might also like her recent collection, Murder By Theory: Two Tales from the Ivory Tower's Dark Side, in which academic murder is explored at greater length.

March 31, 2023 @ 9:42 am · Filed by under Language and science


3 Comments

  1. AntC said,

    March 31, 2023 @ 8:51 pm

    If the measurements are insufficient and/or the model is incomplete, one obtains multiple "degenerate" solutions to the problem.

    Is this different in principle to the Schrödinger's Cat experiment? Why are subatomic Physicists so obsessed with evading guilt for dead mammals? It's ok providing it remains a thought experiment?

  2. John P. said,

    April 3, 2023 @ 6:43 pm

    A humorous April fools paper (I'm not even an academic.)

    I skimmed it mainly because of the strong resemblance to the Bifrost system used in the anime Psycho-Pass (season 3). The author's paper seems to imply more solo planning and execution. Implementing all those degenerate solutions in reality neglects the outsized agency required to execute it, as was seen from the Bifrost system. This actually gives weight to worries over the possibility of prompts by humans given to AI that have access to other human agents (taskrabbit), in that the AI might be able to easy create those degenerate solutions and execute them anonymously over the dark web (the outsized agency is outsourced to AI+human agents) so that no one is discernibly guilty of crimes such as murder.

    The results of the paper (Did the murder succeed? Did the author escape criminal investigation and conviction?) are unreported (literally: "See footnote"). The author ironically complains in the footnote there are too many null results that go unreported, and so for this mystery I can only hypothesize:

    'The author doesn't report the result for fear of incrimination (the author is blissful)', or
    'The author is disappointed that they didn't actually succeed in the murder (the author is devastated and embarrassed),' or
    'The author doesn't report the result for fear of incrimination (the author is agonized over not being able to take credit and share their juicy secret)'.

  3. JPL said,

    April 8, 2023 @ 7:55 pm

    I haven't had time to comment these days, but I just wanted to register, before this post disappears below the fold, that, in spite of the seeming lack of interest here, measured by the number of comments, I, for one, would much appreciate any further posts concerning the work of Dr Armstrong, since, although the article focused above appears to be whimsical in intent, I think her approach to the description of the physical (i.e., natural) world, and that of her fields of expertise, but especially her unique combination of research interests, could be quite useful, not just for phonetics, but also for semantics of natural languages, because it could perhaps provide an alternative interpretive vocabulary for describing the objective situations referred to by object language expressions, similar to the terms used by J J Gibson for describing the (intentional) objects of perceptual activity. The general problem is the relation between logical form and logical [sic] content: what is the origin of "logical content" in the understanding of physical phenomena, and do we ever find in the natural world logical form completely disembodied from something that it is the logical form "of"? (Chomsky's notion of "autonomous syntax" would seem to assume that there is such a situation, but is my doubting of this on the right track?

RSS feed for comments on this post