Brahe, looking for Kepler

Review of Neural Organization
(M. A. Arbib, P. Érdi and J. Szentágothai, MIT Press, 1998)

Shimon Edelman
Dept. of Psychology, 232 Uris Hall
Cornell University, Ithaca, NY 14853-7601
http://kybele.psych.cornell.edu/~edelman

...inkhorns, arrows, loaves, cruses, fetters, axes, trees, bridges, babes in a bathtub, shells, wallets, shears, keys, dragons, lilies, buckshot, beards, hogs, lamps, bellows, beehives, soup-ladles, stars, snakes, anvils, boxes of vaseline, bells, crutches, forceps, stags' horns, watertight boots, hawks, millstones, eyes on a dish, wax candles, aspergills, unicorns.

J. Joyce, Ulysses, p.325
(London: Bodley Head, 1937)

Abstract. This book should be a required reading for any serious student of the brain. The scope and the accessibility of its presentation of the neurobiological data (especially the functional anatomy of select parts of the central nervous system) more than make up for the peculiarities of the theoretical stance it adopts.

A scrambled mosaic

One bright June day in graduate school, I asked my advisor to recommend some reading material for the approaching summer. One of the articles thus recommended left me with an impression sufficiently vivid to prompt total recall on the slightest provocation (such as the groping for an opening line for the present review). The article in question (Gilbert, 1983), albeit informative and well-written, evoked a lingering feeling of disappointment, no doubt because its title -- Microcircuitry of the visual cortex -- had initially sounded misleadingly suggestive to a literally-minded ex-electrical engineer such as myself. That article listed all manner of neurons and their distribution throughout the visual cortex, but, alas, did not quite specify wiring diagrams.

Satisfyingly, Neural Organization (henceforth NeO) does provide some actual wiring diagrams, and not just for the visual cortex. The anatomy of other neocortical, archicortical, thalamic and cerebellar structures is discussed, as well as their development, and models of their function. All this adds up to a grand tableau resembling nothing as much as a Joycean procession of saints ``bearing symbols of their efficacies,'' complete with buckshot, beards, bellows and beehives (not to mention eyes on a dish).¹

The attempt undertaken by the authors of NeO introduce some order into the masses of currently available findings on brain function is highly commendable. That the result appears less orderly than hoped for can be attributed in part to the sheer scale of the enterprise. It is tempting to compare the present state of knowledge in neurobiology (and, on a different level, in cognitive psychology) to that of pre-Ptolemean astronomy: there are observations to be explained, but theories (even wrong ones) are hard to come by. This analogy, however, is strained: the amount of neurobiological data to be explained is vaster than anything that Ptolemy (or even Kepler) had to confront, and the explanation is highly unlikely to consist of a few equations accompanied by a simple diagram. The problem of commenting on books such as NeO, therefore, this: how to evaluate an attempt to reconstruct a huge scrambled mosaic, without having a clear idea of what the big picture is supposed to show, or, for that matter, whether or not there is a single big picture at all.²

In lieu of speculations concerning the existence and the possible nature of the big picture, NeO offers three threads common to all the chapters: structure, function, and dynamics (as reflected in the book's subtitle). I shall comment on these in the reverse order, starting with dynamics.

Dynamics

The argument underlying NeO's construal of dynamics as an all-encompassing explanatory paradigm is that of unity of approach across time scales. The basic premise -- that phenomena involving change over time are properly described by differential equations -- is unimpeachable. I wonder, however, whether it is productive to lump together, say, compartmental models of membrane potential in single neurons on the one hand and models of development of neural wiring on the other hand. Differential equations are ubiquitous in the sciences; most scientists, however, do not take this to mean that all natural phenomena, even in their own field, are merely different aspects of a single whole.³

Fortunately (I think), NeO does not actually adhere to this extreme approach: apart from chapter 4 (which offers a very good survey of ``dynamical system theory'') and a few paragraphs in the last chapter, ``dynamics'' does not play too prominent a role in the book. Typically, dynamics and differential equations tend to crop up where explanatory (rather than descriptive) approaches are as yet unavailable. For example, towards the end of chapter 4, ``invariant'' pattern recognition is finessed into the ``binding problem'' and is offered a solution in the form of von der Malsburg's Dynamic Link Architecture (p.102) -- a model which only becomes relevant if one accepts the prior assumption that binding is indeed a problem.⁴ Likewise, in chapter 8 (which is devoted to the cerebral cortex), one finds dynamical models of ocular dominance formation (nothing is said about why there are ocular dominance columns in the first place), and of thalamo-cortical oscillations (another phenomenon whose function remains obscure).

Function and schemata

If we relegate dynamics to the status of a mathematical means rather than an explanatory end, the issue of function is brought to the fore. Regrettably, NeO takes ``function'' consistently to mean ``how does this bit of the brain function?'' rather than ``what does this bit of the brain do for living?'' In other words, no clear distinction is made between explanation of operation and explanation of goals and means.

The dearth of this latter kind of explanation is apparent throughout the book. For example, the most succinct and explicit description of the function of the visual cortex in all of chapter 8 is found in a quote from J. Maunsell, suggesting that ``while the early stages of processing in the visual pathway provide a faithful representation of the retinal image, later stages of processing in the visual cortex hold representations that emphasize the viewer's current interest'' (p.223). To be sure; but what are these representations, and why do they have the properties they do, and how do they support visual behavior?

The preference for operational rather than functional explanation that one finds in NeO is explicitly justified in section 11.2.2 ( Brain Theory): ``The issue for the brain theorist, then, is to map complex functions, behaviors, and patterns of thought either on the interactions of these rather large entities (anatomically defined brain regions) or on these very small and numerous components (the neurons)'' (p.338). It seems to me that such a mapping would constitute an incomplete theory of the brain, unless it includes an explanation of its function (over and above its operation). The theoretical concepts which NeO brings to bear on this issue -- dynamics, self-organization, and schemata -- are inherently incapable of filling this lacuna. The reason for this is simple: dynamics, self-organization and schemata all belong to a level of explanation that deals with the operation of a system, rather than its function. In particular, attempting to explain the function of a system by saying that it employs schemata is like explaining how a computer program fulfills its function by saying that it employs subroutines (to make this analogy work better imagine applying it to some really complicated piece of software, such as SABRE, the flight reservation system).

NeO devotes the entire chapter 3 to an explanation of the concept of schema. The overview is excellent and wide-ranging, but it does not convince me that recruiting schemata as an explanatory aid (let alone making the concept of schemata an explanation an sich) can advance a theory that does not otherwise address itself to the functional level. Intuitively, the idea of multiple schemata, agents, subroutines, etc. operating in parallel seems to be a very plausible framework for trying to understand brain and behavior. In itself, however, it is not an explanation, and attempts to present it as such tend to give rise to oxymoronic titles such as ``A Robot that Walks: Emergent Behaviors from a Carefully Evolved Network'' (Brooks, 1989); the italics are mine. In other words, if you want your flock of schemata to cooperate and do something useful, these days you still have to carefully engineer your system, and that takes understanding. Moreover, if you ever succeed to have the schemata evolve without supervision, you'd still want to analyze the emergent behavior to gain understanding of what is going on -- just the problem we have with explaining brain function.

Anatomy

The discussion of structure (that is, anatomy) is truly excellent throughout the book (which is only expected, given J. Szentágothai's contributions to this field over the past half-century). The anatomical data are presented in a lucid form, and are accompanied by outstanding illustrations and a good discussion. I found the chapter on the hippocampus, with its ``systems view'' (p.170), especially illuminating. The problem of integrating current ideas about the function (not merely the operation) of hippocampus in animals and in humans is well-presented. The prospects of linking the idea of cognitive maps with more general notions of memory and other cognitive processes are very intriguing. The possibility of the involvement of the hippocampus in the representation of relational information (discussed on pp.182-4)⁵ is fascinating, in view of the present attempts of brain theorists to address the so-called problem of compositionality (Bienenstock and Geman, 1995). I hope this section will prompt further studies and will generate more discussion in the future.

Summary

The last chapter of NeO presents some wonderful opportunities for a lively debate. To quote just one example, the reader is told that ``Digital computers ...have low adaptability'' (p.340); this must mean that a compartmental simulation of a neural circuit running on a digital computer is equally handicapped. For the sake of brevity, I shall forego all these opportunities,⁶ and proceed to summarize my impression of the book.

I believe that this book is a required reading for any serious student of the brain. Even if one disagrees with the theories NeO propounds, the scope and the accessibility of its presentation of the neurobiological data make it more useful than most textbooks and many reviews.⁷ My guess is that NeO will stimulate many readers to try their own hand at the big mosaic, perhaps playing Theon and Hipparchus to a future Ptolemy -- or maybe even Brahe to someone's Kepler.

References

• Bienenstock, E. and Geman, S. (1995). Compositionality in neural systems. In Arbib, M.~A., editor, The handbook of brain theory and neural networks, pages 223--226. MIT Press.
• Brooks, R. A. (1989). A robot that walks: Emergent behaviors from a carefully evolved network. Neural Computation, 1:253--262.
• Edelman, S. (1999). Representation and recognition in vision. MIT Press, Cambridge, MA.
• Eichenbaum, H., Otto, T., and Cohen, N. J. (1994). Two functional components of the hippocampal memory system. Behavioral and Brain Sciences, 17:449--517.
• Gilbert, C. D. (1983). Microcircuitry of the visual cortex. Annual Rev. Neurosci., 6:217--247.
• Marr, D. (1981). Artificial intelligence: a personal view. In Haugeland, J., editor, Mind Design, chapter 4, pages 129--142. MIT Press, Cambridge, MA.

Footnotes

... dish).¹

I wonder to what extent this is intentional; on p.337, this approach to the study of the brain is characterized as ``nonmonolithic.''

... all.²

Marr, in one of his last papers, expressed concern regarding the prospects of developing comprehensive, analytically explicit (``Type I'') theories of some visual sub-modalities, the alternative being the idea that these may turn out to be describable only as bags of unrelated (``Type II'') tricks (Marr, 1981).

... whole.³

As one should expect, there are exceptions, such as S. Grossberg's recurrent invocation of ``neural dynamics'' as an explanatory device.

... problem.⁴

NeO does not define the binding problem until p.244, where it appears in italics, and is described as ``linking representations of a single object, task, or action widely distributed over many brain regions.'' For an alternative view of binding, see (Edelman, 1999), p.246.

... pp.182-4)⁵

A central reference discussed in the text, (Eichenbaum, Otto and Cohen, 1994), does not appear in the book's bibliography.

... opportunities,⁶

I must mention, however, a disturbing typographical error in three places on p.339, where Descartes' res cogitans is spelled ``res cognitans''.

... reviews.⁷

NeO would have been even more useful if the Web links mentioned in appendix B worked. None of the links I tried allowed me to get much farther than the front page of the project; some links were dead, others pointed back to the front page, and some were password-protected.