Wednesday, January 8, 2014

Ewan Dunbar on smart eyes

Ewan sent me a note a couple of days ago discussing an interesting vision article. It seems that the accepted wisdom is that the retina, all by itself, does a whole lot of what we might think of as high level processing. It does this by incorporating cells that “predict” what ought to be there. Ewan, rightly thought that thinking about what our vision friends do would be useful for priming our thoughts about how to approach linguistic phenomena. Impressionistically, some of this stuff reminds me of some old discussions about instruction vs selection theories of “learning” that played out in immunology in the mid 20th century.  Both Chomsky and Massimo Piatelli Palmerini discussed the immunological history as models for how to think about the structure of FL and how UG gets applied to acquiring Gs.[1] Thx Ewan for bringing this to our attention.

*****

I dropped Norbert a note about this review paper about what happens in the retina that came across my desk via the Twittersphere (thanks to DK). He suggested putting something on the blog about it. As I told him, my initial excitement for this paper faded in the best way possible as I read it (what I could understand of it). The abstract had me expecting high-level descriptions of the functions of low level parts so excitingly insightful as to be almost controversially Marrian, a swift jab at the eyes of anyone who would bristle at asking “what (formally) does this thing do in the abstract?” What it delivered in that regard was better, though, namely, high-level descriptions of the functions of low level parts with so much erudite lab bench detail that it winds up sounding self-evident that these bits do that thing, and, by extension, self-evident that we should and indeed must ask for different levels of description as a part of basic research. In hindsight, I feel like the vision talks I have seen all have this sensible flavor, so maybe this isn’t so surprising.

There is a bit more going on here of interest to general cognitive science people. What they actually thought would be stunning, I think, is the fact that the retina has now been found to do a whole lot of fancy stuff---you know, the back of your eye. Not just that, the back of your dumb little pet salamander’s eye---all by its lonesome, no feedback. For example, apparently the retina detects object motion. Now, (a) this doesn't require first recognizing any little object bits or in any way solving the problem of what “objects” look like; what it really means is “non-background motion”, which doesn’t require knowing anything about anything but motion; and (b) one of the cues to which is “differential motion”. If I move my eye around (as you and your salamander brethren are constantly and incessantly doing, whether you know it or not) then everything is moving, naturally. But certain bits of the scene are moving differently (i.e., in different directions). Those must be things. So here’s a kind of prediction: backgrounds keep moving in the same direction. Now, this isn’t too surprising a prediction given that the reason the background is moving is that your eyes are moving, so one might imagine this “prediction” actually relying on a signal from the motor system driving the eye motion itself (although it actually doesn’t, if I understand their stuff right). Here’s a more interesting kind of prediction: the object also keeps moving in the same direction. Suppose I hide a fish in a moving checkerboard. By that, I mean, there is a moving checkerboard background, and all of a sudden, there is movement in a contrasting direction to the background in the shape of a fish. Apparently, when the “object” stops moving, the motion detecting cells continue to be sensitive at the NEXT expected location of the fish. As they correctly point out, this retinal computation implements prior knowledge of inertia (in fact, the bit about the background too, assuming I understand them right that it doesn’t rely on the motor signal). As a bonus, I would expect that this location information could indeed then go on to give other cells clues about object shapes, although I don’t know if this really happens.

So, here’s a cute polemic: see how easy it is to believe in things that aren't there? All you need is prediction, and even a dumb old retina can do it (I think they do this with salamanders or something). If my retina can hallucinate a whole fish then why people cannot get behind my whole phonological system doing absolute neutralization is beyond me. Here’s another one: some neuro-sentence processing people ask whether certain “surprise” effects are due to predictions or to low-level “pre-activation” of expected words. One needn’t be a psycholinguist to understand this question, transposed to whatever domain. Question from the retina gallery: what’s the difference? The high-level description of any such circuitry is that it’s a “predictor” circuit.

Another general lesson, this one from Norbert: inhibition plays a big role in isolating the relevant factors. Here’s one: direction of motion. Something called a “starburst amacrine cell” is responsible for direction-of-motion feature detection (despite my typical early 1980s American Saturday morning advertising childhood, this really calls to mind some kind of psychoactive sea creature rather than any soft chewy candy). If I understand right, all the direction feature detectors would actually be firing for motion in every direction all the time were it not for these devious little creatures. The detector for motion in direction x is wired to all of the starburst cells BUT the ones sensitive to motion in direction x. Why? Because the starburst cells send an INHIBITORY neurotransmitter. So, by gratuitous (and rather nutty-sounding) use of negation, we construct a direction-x detector. This kind of “negative sensitivity,” as Norbert points out, is sort of like indirect negative evidence in language acquisition. To have indirect negative evidence you need quite a bit of background, for it makes no sense to look for what isn't there unless you have some idea of what should be there. So given a fixed set of options, defined by the feature space, and a set of expectations, you can use absence to provide useful information. Input disposes, not proposes. Now, that’s more Norbert talking than me, as I am notorious for squinting at the difference when it comes to language acquisition. But surely it is a useful case to have kicking around to scrutinize the difference, no? The idea being, I take it, that there is something about wiring up ALL BUT x that speaks to a greater level of “prior expectation” than if there was just a “pro-x” detector, one of zillions of positive detectors for everything under the sun. I am inclined to squint because that is the natural position of my eyes, but I can’t help but feel there’s something there. What, I’m not exactly sure. Maybe it’s an interesting case to think about…



[1] See, for example, Massimo’s: "The Rise of Selective Theories: a Case Study and Some Lessons from Immunology". In: Language Learning and Concept Acquisition: Foundational Issues, (A. Marras & W. Demopoulos Eds.), Ablex Publishing Co., Norwood, NJ, pp.117-130 (1987). As I am currently stuck in Montreal waiting for a flight delayed from yesterday, I cannot give you a Chomsky reference. But if I recall correctly, there is some discussion of this in Reflections.

No comments:

Post a Comment