Calvin, W. H. (1991). Islands in the mind: dynamic subdivisions of association cortex and the
emergence of a Darwin Machine. Seminars in the Neurosciences 3 (5)
423-433.
copyright ©1991 by William H. Calvin and Saunders
Webbed Reprint CollectionWilliam H. Calvin University of Washington Box 351800 Seattle WA 98195-1800 USA Email || Home Page || publication list |
William H.
Calvin
"Islands in the Mind: Dynamic Subdivisions of Association Cortex and the Emergence of a Darwin Machine" as it appeared in copyright ©1991 by William
H. Calvin and Saunders |
To model cognitive processing, language construction, and "intelligence" at a neurophysiological level using darwinian evolutionary mechanisms requires more than a survival-of-the-fittest principle. Darwinism is all about the copying success of patterns (typically DNA strings); here I outline a seconds-to-minutes competition between different spatiotemporal firing patterns in a multifunctional cortical workspace. The proposed mechanism for recall from a passive distributed memory into an active working memory is analogous to genotypes and phenotypes. The ephemeral working patterns copy themselves in the manner of wallpaper pattern repeats; they occupy flexible islands in the workspace (useful for multi-tasking and analogical reasoning) that compete with one another for the limited workspace, with a widespread pattern signaling object identification or readiness to act. Pattern evolution is accelerated by cortical equivalents of the roles played by climate change and lowered sea level in island biogeography. Chimeric islands containing a pastiche of patterns are judged against episodic memories in a way that bears some correspondence to the known organization of human language cortex.
One approach to this multifaceted problem is to examine darwinian evolutionary mechanisms, especially as exemplified by island biogeography, for features that known brain mechanisms might mimic. Darwinism is our baseline mechanism for creativity in nature; there may be others, but we know that darwinism's environment-dependent differential reproduction is capable of achieving new stable levels of complexity, shaping up a new species within millennia and, within a week or two, an antibody to fit a novel antigen. Here I propose a seconds-to-minutes neural mechanism that could exhibit the same wide variations-on-a-theme as seen elsewhere in nature, shaped up by a remembered environment, evolving via a pseudo-reproductive process in a population of neural-encoded patterns to yield novel complexities such as this sentence.
One key feature of darwinism emerged only a century after Darwin's original formulation: at bottom, it is all about one-dimensional patterns and their relative reproductive success. In both species evolution and the immune response, evolution is selecting for sequences of nucleotides via the relative reproductive success of the three-dimensional bodies and antibodies created by the one-dimensional patterns. Darwinism's differential reproduction has also been tentatively extended to memes such as word phrases and musical phrases (slogans and advertising jingles sometimes clone themselves in your unwilling brain). There is a population of patterns, and sometimes one particular pattern has a surge of success. Patterns compete and evolve.
Darwinism is usually thought to require a population of individuals but you can still see pattern competition at work in situations where the unitary individual is hard to define, as in corals and grasses (there are clumps of crabgrass and bluegrass, each of which has a characteristic pattern in its DNA, competing for territory in my back yard). Might some spatial or temporal patterns in our brains also be second-to-minute participants in a contest where copying success is important?
Long-term memory, to judge from its integrity following severe disruptions of the second-to-second activities of neurons, is maintained by a spatial pattern of connectivity: synaptic strengths and neuronal branching patterns. A map of the synaptic strengths in the hidden layer of a neurallike network that has been trained to recognize an object serves as a simplified example of the spatial pattern of a long-term memory; while Aristotle thought that mental images were like icons, the map turns out to look nothing like the object itself. Yet it serves to represent it in a manner analogous to the patch of black bars on a grocery package serving to code for the item within. One pattern represents apple, another orange. In the case of a verb-like movement schema such as bring or fling, the patch's pattern is potentially the movement command sequence, analogous to the way that the spatial pattern of a player-piano roll, when scanned, produces the 88-channel movement sequence that generates the melody.
The lack of pigeonholes, inferred from the difficulty of detecting discrete deficits following small lesions, suggests distributed storage. Though a hologram might also suffice, I am going to illustrate distributed memory by using patterns that simply repeat over a substantial region of cortex, in the manner of wallpaper designs. I imagine these repeats arising during learning, e.g., studies of cortical blood flow show that, in the hour-long process of acquiring skill in the performance of a difficult sensorimotor task, the involved regions of cortex are initially large but shrink with skill acquisition.
Working memory might involve the creation of a spatiotemporal pattern of active neuronal firings from the passive spatial-only connectivity pattern; this distinction is something like the phenotype-genotype distinction in population biology. The active pattern resonates against an underlying long-term synaptic pattern, rather as a vehicle's suspension resonates against the ruts in a washboarded road surface. The active pattern is local (a patch or island) and the passive pattern is more widespread (a region), presumably because the active pattern was once widespread. The lifetime of the active pattern corresponds, in this theory, to short-term memory; the foregoing assumptions seem consistent with the essential role of an active process for memory consolidation.
The same region of cortex probably exhibits many preferred modes of oscillation, each corresponding to a discrete memory item (Figure 1); in the wallpaper analogy, the active pattern might resonate with any of the underlying layers of old wallpaper. Shifts in attention, like presets on an organ keyboard, presumably narrow the resonance possibilities to subsets of the many possible patterns, just as neuromodulators serve to functionally rewire the crab's stomatogastric ganglion to produce a different rhythm from the same set of neurons. For present purposes, patches of association cortex are also considered to be multifunctional in the distributed long-term memory sense of Hughlings Jackson and of Lashley -- but also in Kinsbourne's more active sense of second-to-second reassignibility ("When wide areas of the [cortex] are involved in one mental operation... [they] can be used either for a wide-ranging but shallow encoding, or for a single but difficult mental operation"). The existence of local specialties (such as Penfield's experiential responses and Ojemann's isolated spots related to grammar) presumably does not preclude expert areas from assisting in nonexpert tasks, just as neurosurgeons can also function as general medical practitioners.
In analogy to the cache memories of modern computers, the active patterns of working memory might have preferential access to conscious and unconscious mental processing, e.g., new schemata might be compared to the working pattern "phenotypes" rather than to the long-term "genotype" repeating patterns. The active version is the key feature needed to allow patterns to compete via copying success and evolve in the manner of species (Figures 2,3). While it is difficult to imagine how static long-term memory patterns might interact, their phenotypes could presumably compete for workspace territory much as my clumps of bluegrass and crabgrass compete for lawn space. And because we already know something about how competing phenotypes give rise to new patterns at the genotype level in biology, there is a body of theory and strategies to import, which might aid our understanding of cortical pattern evolution.
One might expect similar phenomena from the multi-dimensional schemata patterns; a cortical island with several neighbors, each expressing the same high-contrast pattern, might itself begin to express this pattern during times of lowered insulation, if its original pattern was not robust (the fill-in of scotomata serves as an example). Active patterns, as opposed to the passive ones, are likely to enhance spatial induction of pattern via surround inhibition; as in Mach band phenomena (alternating bands of light and dark bars seen near a boundary), this active pattern might deepen the "moats" immediately around a cortical island while promoting takeover by the pattern at slightly longer distances, e.g., on an adjacent island. Some active patterns are undoubtedly better at bridging gaps than others; the lengths of the axons of corticocortical neurons would also be important for spatial induction propensities.
Whereas the reproductive fitness of an animal species is judged against a real environment of predators, pathogens, nesting sites and food availability, the schema pattern is presumably judged against a remembered environment of life experiences and inherited schemata. It is these long-term memories against which the candidate patterns occasionally resonate to heighten the contrast depth of the pattern and enhance its ability to expand its active territory by repeatedly copying the pattern.
The island subdivisions could disappear on occasion, a particularly successful pattern taking over much of the workspace (Figure 4) in the manner of a single wallpaper pattern repeating over many of the rooms of a house. This might correspond to particularly intense recognition phenomena (such as dj vu) or to concentrating attention on a difficult movement plan (as during get set to throw or dive).
In the manner of moats around castles, a sea of inhibition might prevent an expansionistic pattern from invading an adjacent patch. Multi-tasking and analogical reasoning would seem to require that very different patterns be maintained in a few large patches (Figure 5), as in the six items needed for "A is to B as C is to (pick one of three choices)". Recalling a novel list of items (e.g., digit span) also requires compartmentalizing a number of schemata without confusing them. Manufacturing an Acheulian hand-ax requires keeping eight tasks or search images in mind simultaneously (see Savage-Rumbaugh, this issue).
Second, large continental populations buffer change. Selection pressures are usually most pronounced on the margins of the species range, but this takes a long time to affect a large central population, compared to a small island where all individuals may be on the margins.
When sea level lowers in an archipelago, some islands are reconnected to each other and a competition between species ensues; with further lowering, yet another island joins the conglomerate and the winner of the first competition faces a challenger. When sea level rises, diversification can again proceed as islands are reestablished. This staged competition can presumably operate in cerebral cortex; one functional role of some of the rhythms seen in the electroencephalogram could be to create cycles of competition among the islands; varying the level of the "moat of inhibition" is the obvious possibility. Inhibition per se is not required, only an insulating pattern in the background that cannot be readily invaded by island patterns.
Regional climate change operates somewhat differently; the background against which natural selection operates is substantially changed for many islands simultaneously. In the cortical competition, this could be done by making some long-term memories more accessible than others, shifting attention as it were. Random variations in the bar-code-like dynamic patterns might then stumble upon new resonances with the static synaptic patterns that enhance the pattern's reproductive fitness relative to neighboring dynamic patterns, so that subsequent inter-island competitions are now played against new rivals. Cortical climate change might speed up rare-match recognition tasks, as when the best possible facial schema must be located in order to identify a specific person.
Archipelago biogeography, as compared to continental, may also exhibit faster shaping-up of species and a maintenance of greater diversity. This suggests that individuals -- whether through training or inborn abilities -- able to regulate a cortical workspace, by dynamically subdividing it into even smaller-yet-secure islands, might pari passu become quicker in both their categorizations and analogical reasoning. The fourfold increase of cerebral cortex during hominid evolution,, might aid the multiplicity aspect via increasing workspace size but, without smaller islands, might not aid speed significantly.
The most familiar paradigm for ranking candidate sequences is the traveling salesman problem in mathematics. If one thinks of the islands as cities, and the path taken by a traveling salesman as the itinerary, some sequences (two are shown in Figure 6) will be rated highly because of minimizing costs, while others will be expensive or repeat (as in stuttering). The usual way of approaching the problem is to try a number of alternatives, keeping track of their costs, then ranking them.
Another way of handling the sequence-rating problem arises from chunking. There are typically some limits as to how many schemata we can handle at one time; a phone number longer than seven digits may not be remembered ten seconds later. The limited size of the workspace suggested by digit span creates a need for schemata that are, individually, more intricate. Acronyms serve to chunk long organization names; indeed, many a new word (or a new additional meaning for an old word) is simply a shortcut substitute for a long phrase. Though shorter, they represent the longer phrase as a single unit; psychologists speak of the longer phrase as having been chunked, rather as hay is bundled. A chunk, in the present theory, would be a chimera, as in the goat-sheep chimera that appeared on the cover of Nature a few years ago. This new-style island would contain a pastiche of the patterns from several islands of the cache (Figure 7).
Even real islands can aggregate; the western portion of North America is largely a conglomerate of former islands, recognized as such by the characteristic patterns of rock layering being retained across a land area called a terrane, then changing abruptly. If isolated by sea level rise, a collection of terranes can become a chimeric island (e.g., Vancouver Island). A cortical chimera would be an island that maintained many schematic characteristics of each originating element; one can imagine them arising either by concatenation or by a pastiche copying process. This chunking might free up a portion of the active workspace for new pattern formation.
What the chimera represents depends on what the cache islands represented; phonemes could be chunked into words, and words chunked into word phrases. Note that chunking has the possibility of representing sequence, were the chimera scanned in the proper direction. If the cache islands were phoneme pronunciations, the chimera pattern might serve to pronounce an unfamiliar word. If the cache islands were words (Figure 8), the chimera pattern might serve to pronounce a word phrase. If the cache islands were word phrases, the chimera might represent a complicated sentence with embedded phrases and recursion (while this suggests incredibly long patterns, "pointers" to patterns elsewhere might suffice for some components). If the cache island patterns represented the items on a grocery list, selecting among alternative chimeras could be used to plan an efficient route through the aisles of the grocery store.
In simple manifestations of language such as my tourist German, the mere juxtaposition of words tends to convey a simple message, whether or not I get them into a conventional ordering (such as the subject-verb-object of German declarative sentences or the subject-object-verb that signals a dependent clause). Word order can be an important aid to decoding the sentence, but linguists emphasize that syntax is far more complicated. The listener must make a mental model of the possible relationships between words, and relationships between phrases (often nested) -- and not just one model, but a number of possible ones, as in contemplating Bickerton's ambiguous example: The cow with the crumpled horn that Farmer Giles likes, where the listener must guess whether he likes the cow or the horn. Embedded phrases, like introns in DNA sequences, require special handling, the nature of which is poorly understood. Chimeric islands would, however, seem to have appropriate properties for constructing constituent phrases and evaluating a candidate for an entire sentence.
The sequence itself may be irrelevant in some cases, with the travel plan only serving to identify a subset of the possible collections of schemata and rate the combination (some verbal categories, such as multiple senses of the same word, might be subserved by chimeras). Where sequence is relevant, one might expect to find a convention that determines the direction of scan -- probably not the skewers of Figures 7-9, but something analogous to the start and stop triplets used in the genetic code. A pointer to the movement program for taking-a-breath-and-opening-the-mouth could easily serve to demarcate the front end of the sequence; were such a feature embedded in the long-term static patterns, it would also serve to identify where a pattern repeat began.
Suppose that copying errors or concatenated islands generate a novel sequence of movements or words. Novel patterns would need to be maintained by the copying process, as they would be less well anchored to underlying static patterns (as in the apples-and-oranges example). No candidate sequence will have a perfect fit to a memorized episode but some candidates will fit better than others, e.g., all but the last word. Some novel chimeras will be poorly rated because they violate customary word order, others because they would be inappropriate to the situation. And so we can imagine a darwinian competition between candidate sequences, where the better-than-average candidate chimeras reproduce into the workspace of the poorer-than-average chimeras -- and with copying errors that introduce new variants (such as transpositions or near-synonyms; note the new entries in Round 2 at the bottom of Figure 9).
Sometimes other schemata drawn from cache memory might be substituted or interposed in the manner of introns; some well-learned sequences might be readily evoked directly from long-term memory such as stock phrases and modular movements. Over many rounds, each with such noise introduced during pattern copying, a particularly suitable-to-the-situation sequence might be shaped up (and perhaps stored as a static pattern representing a new episodic memory). A sequence good enough for the intended recipient to reliably decode is all that is usually required; optimal rankings (the focus of traveling salesman research) are probably only attained by poets. I have called this shaping process a "Darwin Machine" to emphasize its similarity to species evolution and the immune response,.
Since a traveling salesman scan through the cache archipelago would suffice to define a sequence, why have a chimeric representation of the subset, one that preserves sequence? One reason is speed: an evolving population of sequence-coded individuals is analogous to parallel computation, while the scan is like using serial computing and keeping a list of path ratings and descriptions for eventual comparison. Whereas a computer programmer might think of such a serial scheme, biology tends to hold copying contests instead.
Chimeras are certainly in the biological tradition of polyploidism and recombination: my chimeras, indeed, were inspired by the concept of the eukaryotic cell as an aggregate of simpler organisms (just as my island subdivisions were influenced by the parcellation concept for neural development, and the widespread concurrent processing was influenced by other recent theories for cerebral cortex,,). We should not be surprised to see chimeras in an intracerebral Darwin Machine; indeed, given that brain circuits are considerably less reproductively constrained (low-cost patterns, ephemeral islands, and no individuals) than are whole organisms (whose reproductive costs are often high), brains might have invented something even more sophisticated than sex.
A low rate of formation of cell assemblies has been suggested as a defect that could underlie some of the abnormalities of consciousness. An instability of islands might interfere with sustained attention tasks (a problem for both schizophrenic and manic-depressive patients, associated with middle prefrontal cortex). Premature closure could result from island evolution when takeover occurs without exploring sufficient variants, and procrastination could represent the stalemate case; this points up the importance of systems that might regulate the evolutionary process via boredom levels and frequent climate changes.
Suppose for a moment that Ojemann's periphery is the site of my cache memory, those single-pattern islands evolving away under the influence of long-term memory patterns of individual schemata, e.g., the words of our vocabulary. And that, much as I have diagramed in Figure 8, Ojemann's core is an area specializing in islands that are chimeras of the peripheral islands, that the chimeras serve to preserve sequence, and that these chimeras separately evolve according to their matches to episodic memories. When the same composite pattern is found on enough such islands, perhaps the pattern becomes good enough to act upon -- and so thought is transformed into action.
Chimeric islands might, of course, be found mixed in with single-schema islands, especially in the case of a shortcut word such as an acronym (a sequence that has become a single word). But for purposes of competitions between evolving sequences, concentrating the chimeric islands would allow them to more effectively compete with one another (and allow for a cortical region that specialized in episodic memories). Given the reputation of prefrontal cortex for monitoring narratives and constructing plans, Ojemann's core is surely not the only cosmopolitan region of brain.
A traveling-salesman solution is, however, inadequate for some sequencing tasks such as throwing accurately1,, because of the substantial inherent timing jitter of all known neurons. Fortunately, there is a solution via the Law of Large Numbers (this is the principle utilized by every experimenter trying to halve a standard deviation by using four times as many subjects). Just as the jitter in the interval between heartbeats decreases with the inverse square root of the number of coupled pacemaker cells, so a repeating pattern corresponding to the movement schema for throwing (simultaneously scanned at many repeats and the results averaged) can be used to reduce performance jitter. This would be under considerable natural selection24: throwing twice as far with a comparable hit rate is always better for hunting. But to reliably release the projectile within the narrowed launch window requires reducing timing jitter eight-fold -- and that requires 64 repeats of the throwing schema pattern to be orchestrated as a chorus. (The Law of Large Numbers approach also applies to hyperacuity in precision sensory tasks).
Perhaps the archipelago of chimeric islands came later, allowing a Darwin Machine to help select among alternate ballistic movement schemata (as is handy when throwing in novel situations, rather than from a standard distance), with other spare-time uses of the same neural machinery then developing (multi-tasking, language, long-term planning, analogical reasoning). Both scenarios could be correct: precision tasks might have come under natural selection (expanding the workspace) and then multi-tasking might have dominated brain evolution for awhile (improving subdivisions of the workspace into stable islands), or vice versa. Patterns on quasi-stable islands -- and especially the regulation of their variety and recombination -- would seem a key feature of either evolutionary road to higher intellectual function.
REFERENCES omitted. String search my collected bibliography.