Keith Sutherland's penetrating critique of Thomas Metzinger (JCS posting 21 April) illustrates the confusions one can get into by using pre-computational ideas.
Keith worries that since there are 'top-down' effects (eg in perception) that any kind of 'bottom-up' attempt to to resolve the 'binding problem' in terms of phase-locked oscillations in the feature-detection neurons involved in the perception of a particular object must fail. The central issue, for Keith (and apparently for Hardcastle), is causality:
.... The obvious question that one would ask is whether the oscillations are the cause or the consequence of perceptual binding. Is it just a spurious correlation? Put more directly:Interestingly, such phase-lockings in global patterns of oscillation are the basis of the only plausible account I know (due to Shastri and his students) of how a connectionist network could perform reasoning in real time. Shastri correctly regards the network as a kind of computer, and asks how it can 'run' nontrivial 'software' (scare quotes because these are both rather different than the usual senses in Von Neumann architectures). Now the question that so bothers Keith is meaningless. Does hardware cause software or vice versa? The proper answer is that they are causally intertwined: each is constantly causing changes in the other. (Nervous systems often seem to use temporal phase information: we havnt noticed it until recently partly because the techniques of gathering data from active neurones have made it invisible. Its the only way that bats could possibly react as well as they do to acoustic signals an order of magnitude faster than the firing rate of any single neuron, for example: see Salient, in Tucson II abstracts)
how does the brain "know" that the output of certain feature- detecting neurons is part of any particular object without a prior (top-down) model of that object?Or as Hardcastle puts it "They must somehow work out how to indicate that this neuron over here with this particular receptive field is signalling a bit of the same feature as that neuron over there with that particular receptive field" (p. 77). There is no way that this could be accounted for by synchronous oscillations alone as this would require some prior relational knowledge by the system. If there are other cues (neighbourhood, similarity, continuity of motion, common retinal location etc.) then synchronous oscillations are the effect, not the cause of perceptual binding.
Conventional computer architectures are designed so as to separate these causalites as far as possible for reasons of engineering and commercial convenience, but it isn't easy to do. The results are often called problems of portability or compatibility: basically, a program will only run on one machine since their causal dependency has become too mutual. This is more like the natural state of software, and in neural(ly plausible) architectures it becomes quite thorough.
Computer science provides a framework which can make these ideas precise, fortunately. The missing idea that gives solves Keith's dilemma is that of the virtual machine. To oversimplify somewhat: the 'bottom-up' 40 Hz oscillation is the activity of the 'processor' of this machine: Metzinger's 'top-down' models of objects, scenes and the phenomenal self are constructs in the software of this machine. (Scarequotes again because a connectionist computer in a sense is all processor as well as all memory.) To ask which causes the other, as though they were two separate processes or objects, is just a conceptual error. And is it 'purely bottom-up'? Well, it is if you look only at the hardware. But when the software is running, it is 'top-down', since running software has causal consequences for the hardware. These maps arent "passive", even though the hardware is completely bottom-up and nonmystical. They arent little green men, but they sure make things happen.
This kind of approach, by the way, positively celebrates such things as Necker cubes and Rubin vases: it even can account for things like the minutea of timedelays in switching perceptual gestalts and why one cannot have two at once, etc..
Pat Hayes
Phayes@uiuc.edu
PS. Can't help remarking on this:
[KS] There is a wide-spread belief within the cognitive sciences that the phenomenal self is a constructed illusion. This is partly due to a robust homuphobia but in some cases also seems to result from a misunderstanding of certain teachings of the Buddha. For example, Claxton (1996) makes a remarkably similar case to Metzinger, but acknowledges explicity his inspiration -- Abhidhamma Buddhism. I have no way of knowing whether Metzinger draws on similar sources, but authors in general could be more explicit as to how their own philosophical and religious convictions affect their theory building.This is a wonderful passage, I may use it the next time I teach an introductory class in logic. Keith is SO clever at these rhetorical tricks! Notice how he has (totally spuriously) linked the view he is attacking with religion, frowningly hinted at Metzinger's dishonesty at not admitting to mystic inspiration (without actually accusing him of anything, of course); nodded forgivingly in a high-falutin' kind of way, but then suddenly spat "dogma" at it like a cat's hiss.I don't actually have any a priori objection to incorporating religious insights to help unravel the problem of consciousness. As John Searle pointed out at Tucson, now that we live in a de-mystified world there is no principled problem in including such insights within a testable natural-science approach. But, I would maintain, in the field of human consciousness it is better to start with our own experience rather than dogma.
I suspect that Metzinger, like me, has not the slightest idea or interest in what a mystic sage from centuries ago said or did. Like most clear-thinking people, I plan to go to my grave without reading a word of Abhidhamma Buddhism.
[KS] One's own experience, which is perhaps the only starting point, seems to be centred around a phenomenal experiencer.The key is that little word "seems". Indeed that is what it seems like. So we need a theory which predicts not that it IS centered around such an experiencer, but that experiences will SEEM to be centered in that way.
jcs-online menu