A brain mechanism for recursive embedding (such as sentences within sentences: I think I saw him leave to go home) is considered essential for Universal Grammar (features common to all known languages except pidgins). Among the linguists’ other desiderata are mechanisms for long-range dependencies, including binding of pronouns to their referents. Such binding requires longer-than-local links; recursive embedding requires structuring a hierarchy of them.
     Corticocortical axon bundles are considerably worse than the incoherent fiber optic bundles where neighbors fail to remain neighbors. I have predicted (Soc. Neurosci. Abstr. 1993), however, an error-correction ability that should emerge from the lattice connectivity of the superficial pyramidal neurons. Were it not well tuned, linkages would be restricted to well-practiced special cases, analogous to the mariners’ signal flags — and to only a few at a time, limiting the possible associations conveyed. Embedding would be restricted to stock phrases. Properly tuned up, the error-correction circuitry could reconstitute an arbitrary spatiotemporal firing pattern in the target cortex — and thus convey novel associations, even passing them through a hierarchy of embeddings. This makes corticocortical coherence a candidate for what converted protolanguage (structureless pidgins) into Language Itself (linguists doubt that intermediate forms exist). Indeed, the transition from special-case to arbitrary code conveyance could have implemented both major innovations of Universal Grammar — embedding and long-range links. The versatile linkages described here, in the context of a darwinian process at each end to bootstrap quality, might also have helped convert the infrequently-innovating cultures of Homo erectus into the constantly-evolving ones of Homo sapiens.

     Funding ;-) via my two new books, How Brains Think (BasicBooks 1996) and The Cerebral Code (MIT Press 1996).

     The pyramidal neurons of the superficial layers of neocortex have, in addition to long corticocortical axons that go through the white matter, many collaterals that stay within the superficial layers. These intrinsic horizontal connections are excitatory and constitute about 80% of the input to other superficial pyramidal neurons.

     Unlike the deep-layer pyramidal neurons, the collaterals of superficial pyramidal neurons have an interesting "express train" patterning, with no "stops" except in clusters separated by about 0.5 mm (varies with cortical area). They also have many terminations locally, presumably to other pyramidal neurons in the same dendritic bundle. The overall pattern is like that of a focussed flashlight beam, a hot spot in the middle and then a bright ring about 0.5 mm away.
      The flashlight-beam patterning means that neurons about 0.5 mm apart recurrently excite one another. This makes it likely that they will synchronize their firing, if otherwise active. But the bright rings will also cross at a third and fourth point to form a triangular array of synchronized neurons.
      If many such triangular arrays form, the spatiotemporal firing pattern will exhibit repeats; the minimum replicable unit with no redundancy is a hexagon 0.5 mm across. Thus one can think of clones of movement commands forming a hexagonal mosaic, with alternative movements competing with one another for space in association cortex.

CRYSTALLIZED CODE

      Triangular arrays of synchronized feature detectors not only recruit territory but they standardize within their territory ("error correction"). If the long corticocortical axon branches have the same flashlight-beam fanout at their distant terminus as the short intralaminar branches do back home, a large "sending territory" ought to be capable of recreating a smaller distant territory. If all the triangular arrays in a unit hexagon do this, then the complete spatiotemporal pattern can be recreated at a distance (the "faux fax").

LANGUAGE, Plain and fancy

• Most animal "languages" are merely a set of utterances, each with a conventional meaning. Two-word utterances are rare in nature, possible with training.
• Protolanguage is the fancier (but unstructured) word association seen in children under two, agrammatic aphasics, the tutored apes, and speakers of pidgins. It takes a lot of time to say who did what to whom.
• Universal Grammar (UG) is the set of features common to all known human languages, such as nested embedding (I think I saw him leave to go home), binding (pronouns refer to word phrases in previous sentences), and the long-range dependencies that Mark Twain complained about:

Whenever the literary German dives into a sentence, that is the last you are going to see of him till he emerges on the other side of the Atlantic with his verb in his mouth.

Linguists say there are no primitive languages (say, with no pronouns, or with embedding only one or two deep); it's as if UG structures come as a single big step up from protolanguage. What neural circuitry innovation might tie together pronoun binding with recursive embedding? I have a candidate.

     Working memory in humans is thought to involve spatiotemporal firing patterns, while spatial-only passive representations suffice for fading short-term memory and long-term memory. Either could implement associations (ANNs assume spatial-only patterns in the connectivity).

     Associating two items is easy, both in passive ANNs and in the active superposition of two spatiotemporal firing patterns at a hexagonal boundary. But associations are not always learned through repetition. We can create never-seen-before unique associations such as transparent shoe and communicate this mental model to someone else. This suggests active patterns.
      Long-range corticocortical connections, moreover, can superimpose a number of active spatiotemporal patterns in a convergence zone. One can imagine a spatiotemporal pattern representing

      ANNs show that almost any spatial input pattern can be learned. That's fortunate because most corticocortical paths probably severely distort the pattern sent. The fanout of an axon terminal is equivalent to blur and neighbors failing to remain neighbors (jumble) is similar to what happens in fiber-optic cables when they are longer than those used in endoscopes and similar imaging applications requiring coherence. An ANN-like substitution of one code for another is thus what we expect from the incoherent neuroanatomy. But language is about novel associations -- if paraphrasing, for example, we don't have time to learn each association.
      Furthermore, for structured sentences, we need to maintain an audit trail - and that profits greatly from a common code in all cortical areas, i.e., that the spatiotemporal firing pattern for shoe is the same here and there, so that patterns can be passed through intermediate way-stations undistorted until reaching an area with a sustaining resonance. That's what temporarily coherent corticocorticals allow, thanks to the same error correction that standardizes the pattern locally.

      If the spatiotemporal firing pattern within a hexagon is analogous to a tune, then a hexagonal mosaic is a plainchant choir. The competition for recruits constitutes "dueling choirs." But with coherent corticocorticals, the choir need not be contiguous: it can be distributed, rather like being unable to attend choir practice but contributing to the critical mass by speakerphone conference calls. That means that the choir "singing" a prepositional phrase could contribute to the success of a distant choir representing the whole sentence, in its competition with alternative stories.

      Different melodies can also overlap (say, black with shoe), and a sentence is starting to look like a symphony of many voices rather than a Gregorian chant. While such composite codes must compete with one another via winning territory, we often need to "read out" the winner because of the serial-order bottleneck of speech. This is like a backward version of Benjamin Britten's Young Person's Guide to the Orchestra where the instruments first play separately, then together. Here we read out each voice, one by one, in a conventional order.
      Reading out this deep-structure pattern (the successful "meaning") into surface-structure conventions means unpacking the overlain constituent melodies. That can be achieved by interrogating the mosaics elsewhere that contributed a prepositional phrase or noun phrase ("Who said X? Sing it again, the whole thing!"). This prevents straying associations such as blond black man.

      If corticocortical coherence were to degenerate into the AAN-like learned associations expected from the incoherent neuroanatomy, we would expect to lose binding, nested embedding, and similar features promoted by versatile superposition and audit trails.
      Similarly, a widespread use of hexagonal mosaics could have enabled the step up from protolanguage to UG. Because the cultures of Homo erectus were so unchanging, the transition to rapidly-innovating Homo sapiens may have been facilitated by widespread use of quality-bootstrapping darwinian processes operating on the msec-to-minute time scale.
      One side effect of the darwinian hexagonal mosaics, consistent with the step up to UG, could have been the coherent corticocortical transmission that allowed for a common code in many cortical areas -- and thus rapid structured associations without prior learning.

The corticocortical coherence chapter of THE CEREBRAL CODE and the syntax chapter of HOW BRAINS THINK (BasicBooks 1996) may be of interest.

Language cortex physiology is addressed in various chapters of CONVERSATIONS WITH NEIL'S BRAIN (Addison-Wesley 1994).

Universal grammar is nicely addressed in Ray Jackendoff's book PATTERNS IN THE MIND (BasicBooks 1993), and protolanguage is similarly addressed in Derek Bickerton's LANGUAGE AND SPECIES (University of Chicago Press 1990).

For more related reading, see the Calvin Bookshelf.