I find Metzinger's article a mixture of vague philosophizing based on highly selective empirical data. He maintains his bottom-up viewpoint -- mit Gewalt, as we say in German -- by totally disregarding cortical anatomy that shows massive neural pathways going from every higher cortical level back to those lower regions from which it received its information in the first place. The conclusion is unavoidable: The data that are fed up the cognitive ladder, are filtered, modified, added to and censored by higher structures. If that is not top-down control, then we have a profound semantic problem.
But everything goes up in Metzinger's model until it winds up as a "holistic entity of the highest order". The author then speaks of the "single psychological moment". I believe he refers here to a common fallacy which led Jeffrey Gray at Tucson II to conclude that consciousness cannot be of any adaptive advantage. He cited, as those present will recall, the tennis player who cannot wait for consciousness to tell him how to hit the ball, believing that consciousness is forever limping behind reality, hence unable to affect current behavior. It strikes me that this views consciousness as a bullet traversing the phase space of brain dynamics. At any given moment the bullet is somewhere, that is to say, the "psychologicaI moment" is rife with holistic entities. I would argue, on the contrary, that -- if I could stop all brain activity -- there would be no content, no psychological moment. It appears to me that one of the outstanding characteristics of consciousness is the fact that it straddles time. Hundreds of milliseconds before the ball arrives at his court, the player became aware of the location and movement of the opponent when he struck the ball. With plenty of time to compute ahead, the player not only knows how and when the ball will arrive, but, looking into the future, he has already prepared for his return stroke. Late arriving sensations will produce reflexive midcourse corrections. I see most of sensory data already obsolete by the time they arrive at the cortex, anticipated, that is, by consciousness.
But this is an aside. Metzinger now goes into a tedious discussion of International Klein Blue that is to illustrate what he calls the homogeneity of perception. Again, I would argue for the opposite. It is my (subjective) impression that even elementary sensations like International Klein Blue, will pull in associations such as Metzinger's book cover or a recent ocean voyage. They may not all reach consciousness, as I will explain below.
I do agree with Metzinger on his definition of the binding problem and his assertion that, in addition, the superposition problem will have to be solved. But, not his way. The questionable binding supposedly achieved by the simultaneous firing of many neurons, and by the now famous 40 hz rhythm, has always struck me as a desperate way out of a dilemma. If we needed a homunculus to view the spatial character of neural activity, who or what in the brain is the necessary timekeeper? Simultaneity of neural firing means something to the physiologist who has placed electrodes in two or more neurons and compares outputs on an oscilloscope screen. But time is an ill-defined quantity within a system that is multiply connected along paths of different lengths, and without a unique velocity of signal propagation. I might point out in passing, that I discovered a distinct and striking periodic response to photic stimuli in frog retinal ganglion cells (see M. Stiles et al., Experimental Neurology 88, 176-197, 1985). The oscillations were in the range of 15 to 50 hz. Must we conclude that frogs are conscious even if anesthetized and curarized?
Metzinger solves the temporal homunculus problem by postulating "coherence detectors at higher levels." Just how they would function to tell us what we are looking at, is cloaked in such obtuse language as "higher order embedding relations." Also, if such high order coherence detectors exist, they must also carry the full complement of information concerning all features that are being so united, otherwise they would only tell us that something is being looked at. We would need labeled sensory convergence zones, which have, in fact, been looked for intensively, but never found.
I want to talk briefly now about my own work that addresses the same problems, but provides answers with a lot less hokus-pokus. I am referring to what I have called the sketchpad model. It has been published in various forms [Science 237, 184-187 (1987); Consciousness & Cognition 4, 346-368 (1995); in Toward a Science of Consciousness, Hameroff et al. Eds., MIT Press 1996 among others]
The sketchpad model is based on these empirically known facts. (I use the visual system as illustration of the phenomena observed.): Massive recurrent neural pathways are directed from higher to lower cortical centers and from cortex to thalamic relay nuclei. Mental images are formed at peripheral sensory centers where primary sensory data are normally displayed in retinotopic form. (See, for example, the work of S. Kosslyn.) Mental images can act as quasi-sensory input to higher centers (Kosslyn, again).
I add the following plausible assumptions: The actual input to higher centers consists of a concatenation of raw sensory data and modifications controlled from above through the corticofugal pathways. Thus, a competition exists at these peripheral sensory levels between raw sense data and cortical fancy; In the absence of sensory input, the mental images are pure fancy.
The top-down control of peripheral images thus requires an inversion of normal, bottom-up sensory processing. The problem of the inverse is solved in my model by assuming the action of hill-climbing, or optimization processes for which I have suggested simple neural mechanisms involving LGN, VI, nucleus reticularis thalami (NRT), and the ascending brainstem reticular formation (see e.g. the article in Consc. & Cogni.). The cost function to be optimized is the output of cells of the reticular activating system that are an integral representation of cortical activity, compressed into a simple scalar quantity. It is known that these cells are inhibitory on neurons in the NRT, which, in turn, are inhibitory on thalamic relay cells and thus control transmission of visual data to VI in the cortex. The optimization process, which I have located in the thalamic relay nuclei, but which similarly may occur at higher levels, will tend to generate or reinforce, at the level of the LGN, those stimulus patterns that maximize (or minimize) the brainstem input. The mechanisms have been demonstrated in numerous computer simulation studies.
Let me illustrate the process with the following thought experiment. Let us start with something like a free association test in psychology, except the subject is not asked to report his/her associations. He/she is given the word ICE. This will activate a number of memory traces, or pull them out of long-term storage into a region (perhaps in the prefrontal cortex) that has been called working (or short term) memory. Among these may be the memory of icicles hanging from the roof, or of falling on ice, or a frozen lake, or ice cubes jingling in a whiskey glass. Any one or all of these may be simultaneously present in working memory, but consciousness is a narrow channel that is selective and exclusive. The different associations will now compete for entrance into consciousness. Only one at a time can succeed (by hill-climbing and top-down control) in generating the appropriate images in the various sensory modalities. The nascent images further strengthen the successful traces in working memory in this cyclic, self-referent process. If the images are of falling on ice, the memory of a broken limb may ensue, and the stream-of-consciousness may go on from there. You may never realize the fleeting (unconscious) association with whiskey glasses.
The mechanism of image modification by top-down control has the following desirable consequences:
Binding. The various sensory features of an event are brought together through the process of optimization that generates peripherally an image of the event, that is either reinforced sensory or confabulated. Pattern completion of indistinct or incomplete sense data is part of the same process.
The superposition problem is solved by the selectivity and exclusivity of the process (This is achieved in the simulation experiments by making the cost function a non-linear superposition of individual cortical feature analyzers). On the other hand, the process promotes the simultaneous expression of features in different modalities belonging to the same physical reality (the visual image of the whiskey glass and the jangling of the ice cubes).
Homunculus. This specter can finally be laid to rest. The images that are assembled in peripheral sensory areas (perhaps as distal as the thalamus) in what represents a Cartesian theater of sorts, are looked at, judged, and further elaborated by the higher levels of the brain. This differs fundamentally from most other theories that search for a final meaningful output state. The process is, I emphasize again, cyclic, selfreferent, with the binding occurring at the bottom rather than the top of the sensory ladder.
Sketchpad. The model allows for a process that I feel is the prime function of consciousness. We try out scenarios not by actually performing them, but by creating mental images that we can then judge, evaluate, and elaborate. Any creative human act has that character. This is why I sometimes call the process a sketchpad or a creative loop.
Erich Harth
harth@suhep.phy.syr.edu
jcs-online menu