Neural Net Processing: A Key Factor in the Evolution of Language

© Stephen Plowright
MacKaos Consulting.


This paper is an attempt to examine the implications of recent research in neurocomputing for the question of the evolution of language. An overview of the existing research is given. A model for the evolution of language is constructed which postulates peripatric speciation giving rise to a chance enhancement of vocal articulation, followed by a rapid development of a generative language with a corresponding adaptation for speech. Language may have evolved very rapidly driving the selection for the features of modern anatomy. Also, the development of language was unique to this species and the major factor in its success.


The question of the evolution of language is necessarily speculative, and has indeed been a matter of considerable controversy since it was first debated. Research has typically concentrated on drawing inferences from the evidence of artefacts and biology of early humans, and the capabilities and biology of modern humans. Some interpretations have lacked a robust logical foundation, as has been pointed out by Davidson and Noble (1989, 1993) and Lieberman et al (1992).

Arguments which present features of language acquisition in modern children as evidence for similar stages in the evolution of language in the species as a whole are clearly fallacious. A modern child with a known aptitude for language acquisition, in a language intensive cultural context, is in no way comparable to an evolving species with no linguistic context.

Also cited as fallacious, are arguments which present the existence of isolated anatomical features as evidence of language. It is not unusual for species to evolve structures for a particular purpose, which are later adapted for other functions. Falk (1989), in reply to Davidson and Noble (1989), states with complete conviction that "human language depends on a lateralised brain". This statement, seen by some as an inviolable assumption, does not entirely accord with the evidence of neurophysiology. It is known that, following early left hemisphere damage, the right hemisphere is quite capable of learning language (Feldman, 1992). Also, examples are not uncommon, up to 5%, of normal individuals who exhibit language processing in the right hemisphere, or even distributed bilaterally (Cook, 1986; Corballis, 1991). Lateralisation, in itself, is neither evidence of, nor a prerequisite for the production of language.

The sophistication of artefacts or human achievements can, as cited by Davidson & Noble (1989), offer evidence for the latest date by which language must have been in use. They cannot offer much to the evidence of the earliest date of language use.

Studies with Chimpanzees, it could be argued, reveal more about the likely limitations of early Hominid communication than they do about the linguistic abilities of our early ancestors.

This paper will attempt to bring together the results of recent research in various fields, notably properties of adaptive neural net information processing, to arrive at what is bound to be an initially unpopular view of human and linguistic evolution.


The Fossil Record

The interpretation of fossil evidence is another area of considerable controversy. The following is a superficial overview of the facts which find general agreement at present.

The dominant trend in human evolution has been one of increasing adaptability and brain size. While the various Australopithecines, and probably early Homo, occupied different ecological niches, later Homo was increasingly able to occupy all of these niches. The disadvantage of the less specialised morphology in competition with more specialised species was countered with increased brain power allowing more effective adaptive strategies.

Emerging as sole survivor at some time after 2 myr in Africa and Asia (Swisher et al, 1994) were variants of Homo erectus . This species evolved slowly for 1.5 million years, it demonstrated basic improvements in stone tools and probably used fire.

After 1/2 myr variant Homo forms started to emerge, culminating in Neanderthal and archaic Homo sapiens types by 100,000 B.P. The number of different forms and their relationships to each other, or to modern humans, are subjects of debate at present.

At around 100,000 B.P. anatomically modern humans emerged. The earliest examples were found at Qafzeh, in Israel, some dated at over 90,000 B.P. This species, Homo sapiens in its anatomically modern form, had replaced all other Homo types by 35,000 B.P.

The details of the fossil record are controversial, but the feature which is pertinent to this paper is the rapidity and nature of the morphological changes of the final transition to anatomically modern humans.

Neural Nets

Modern computers have made it possible to simulate some of the processes observed in the brain. In particular, recent work done with adaptive neural nets has produced results which shed much light on the observed features of brain development.
Bearing in mind that any simulation will represent a very small number of neural components compared to real brain systems, this research is not expected to reproduce the complexities of human or animal behaviour. What these simulations do reveal is the way in which the brain constructs complex computational systems in response to information from its environment.

In the simplest case, a neural net is an array of elements in which each element has inputs from other elements, see Figure 1. Each element or neuron sums the inputs (Xi) and produces a signal on its outputs (Zi) if the sum of the inputs is greater than a threshold value. Each of the neuronÕs inputs is weighted according to its input history. There are a number of ways the weight vectors (Wi) can be altered to produce an adaptive net (Hecht-Nielsen, 1990; Hertz et al, 1990).

Figure 1. Neurons with weight vectors Wi = (Wi1, Wi2, .... Win).

Researchers have constructed nets to learn many tasks. The common finding is that, although these nets are not explicitly programmed with information or learning strategies, they will categorise signals according to features of the input. Pattern or feature recognition arises as a result of the self organising nature of these nets. This kind of processing does indeed seem to be the dominant type of processing in biological neural systems.

The advantage of neural processing is that information processing occurs as a wave of activity across a large array of elements allowing sophisticated operations to happen in a small number of steps, typically under 100. As the wave of activation propagates through the net, features of the input trigger a pattern of firing in elements of the next layer, which in turn respond to the pattern in a unique way. In this way, a large number of features are grouped into categories, which in turn converge to be placed into a final category or a unique solution. If a unique solution is not found, the most appropriate category is found. This is often termed a 'parallel distributive process', and the large data structures arising from the categorisation process are known as "schemas". These schemas allow the brain to recognise a wide variety of new information by interpreting through interpolation (Rumelhart, 1989).

The remarkable property of these adaptive nets is that, without being programmed with any knowledge or strategy, and by virtue of their structure alone, they are able to learn quickly to organise information into categories, effectively creating topological feature maps of their information environment (Hertz et al, 1990; Allinson & Johnson, 1989).

The Brain

Much has been made of the importance of lateral asymmetry in the brain. It should be noted that the brain shows more bilateral symmetry than any other major internal organ except the kidneys (Cook, 1986). Although the left hemisphere may be subtly better adapted for language, given the extreme plasticity of the developing brain, and the ability of the right hemisphere to acquire language after early damage to the left, the observed difference between hemispheres is more likely to be one of acquired function, not one of gross structure. This is also supported by the findings of Feldman et al (1992) that both left and right unilateral brain damage will cause initial delays in language development followed by normal rates of progress with no significant differences between the two groups.

Furthermore, proof of the existence of gross structural features from Hominid endocasts contributes nothing to the evidence of Hominid linguistic ability. These features may have been adapted for other functions before the development of language.

As Gerald Edelman points out in Neural Darwinism (1987:143):

"Inasmuch as the functions of a given component of the brain are not necessarily locally defined, such evolutionary deductions from comparative neuroanatomy . . . must be based mainly on morphological and structural criteria; that is, homology can be inferred only for anatomic structures and not for function."

The advantage of lateralisation is not that it allows the acquisition of language, it is that various functions can be organised in the most efficient way. The left hemisphere would not have been wasted in pre-language Homo, it would have been used to organise other features of the information environment such as animal behaviour and prediction of predator/prey movements, hunting strategies, tool making, and social interaction.

A possible mechanism for lateralisation is a differential growth schedule for different areas of the brain. The language areas are maturing at the time of language acquisition, while the contralateral areas mature a little later when spatial skills are learned (Corballis, 1991). Such a process would have survival value as it would effectively lengthen the period of maximum plasticity allowing more cognitive skills to be acquired.

If the brain is to be used as evidence, it must be in terms of its general information processing properties. Recent research in neurophysiology and neurocomputing has provided convergent models of neural processing which point toward categorisation and feature mapping as the key to understanding the power and versatility of the brain. Studies with pigeons and pre-language human infants show a remarkable ability to categorise and generalise without language and without tuition (Edelman, 1987).

The Supralaryngeal Vocal Tract (SVT)

The most reasoned and convincing research on the physiology of speech production in the evolutionary context is that of Lieberman et al (1992). Their work throws much doubt on the likelihood of language in pre-modern and Neandertal humans. Although they find that SVTs of non-modern human type may have been able to produce many of the sounds of speech, they were not well adapted for that purpose.

It is important to consider that rapid articulate speech requires fine motor control not available to other primates. It seems unlikely that such control would have evolved without noticeable adaptation of the SVT for speech.

Figure 2. The anatomy of speech production (after Lieberman et al, 1992).

Duchin (1990) compares metrics of the oral cavities of Pan troglodytes, H. erectus, H. sapiens neanderthalensis, and modern humans concluding that H. erectus, and H. sapiens neanderthalensis, were capable of articulate speech. Leiberman et al (1992), however, point out the importance of the lowered larynx, unique to modern anatomy, in the production of rapidly changing formant frequencies which characterise human speech and which make rapid and accurate encoding and decoding possible. Note the length of the pharynx is similar to the length of the oral cavity, see Figure 2. Such a configuration was not possible in other species.


While many species are known to communicate, only humans are known to be capable of language. The difference is in the generative nature of language. Communication can relate facts about the immediate environment, or the state of an individual, but only with language can we relate things which are completely new.

Only a language speaker could ask "Tell me something that never happens" and only a language speaker could reply to such a request with an original and meaningful response, eg "A dog flying". A four year old human would have no trouble responding appropriately, we are yet to see any other species master such a task. This is not the same as simple lying, in which an incorrect response is given for some chance of gain. Simple lying is not generative as the context is set by the situation or by the querant.

All languages have the same basic hierarchical order: a perceived number of discrete units of sound, or phonemes, which are a subset of the possible human sounds. Clusters of phonemes with corresponding meanings, these are words and morphemes. Concatenation of words to form sentences, no language is restricted to single word discourse. Rules of sentence structure, no language concatenates in a random way (Kess, 1990).

This hierarchical structure is not surprising in the light of neural processing, it is exactly what would be expected. One need not postulate a special type of circuitry, the order seen in language is a natural product of the way neural nets process information. Lakoff (1987) proposes that language and thought are fundamentally structured by spatial metaphor. Lakoff sees spatial metaphorical mapping and transformations as the general way humans understand abstract concepts (Jubak, 1992). Perhaps the difference between the two hemispheres is not as great as first thought. While the nature of verbal and spatial information differ in form and modality, they are both processed in much the same manner.
The prerequisites for language are, firstly, an efficient channel. This requires an organ which is capable of a sufficient range and speed of articulation, and is easily and accurately sensed by the receiver.

Secondly, sufficient processing power to recognise and categorise a large number of linguistic units and clusters of units, along with the rules of grammar.

Thirdly, sufficient short term memory to use as a buffer to hold a meaningful amount, a sentence, of the received message for processing. If the buffer is too small, or the articulation too slow, grammatical structure can not evolve.

A Model for Human Evolution

Middle Pleistocene Homo had a fairly large brain and was able to adapt to a variety of environments. Although they made tools and spread across large areas of the globe, there is no evidence of any sea crossings, nor other co-operative ventures of this type. Also, biological adaptation to verbal language cannot be demonstrated.

There probably was some form of gestural and vocal communication but it is unlikely to have gone beyond pointing, motioning and imitation, like the hunting signs used by some hunter-gatherers today. Most of the gesturing behaviour would have been directed toward food gathering and survival. It is unlikely that most infants would have been exposed to a sufficient number of discrete gestures often enough, or rapidly enough, to learn it as a generative language. Also, if gesture had become generative, it would have conferred such a survival advantage that there would have been rapid selection for adaptations to gestural language. Verbal communication would have conferred more disadvantages than advantages over gesture and would never have become dominant. Although deaf people can acquire signing as a first language, it must be remembered that these sign languages are derived from verbal languages, there is no known example of a fully generative native sign language.

At some time probably around 100,000 B.P., probably in Africa (Cavalli-Sforza, 1989; Stringer, 1984), a group from one of the varieties of Homo became isolated. Inbreeding in a species already undergoing change produced some features which would have reduced its probability of survival, except that one feature conferred an advantage which more than compensated. Such peripatric speciation can result from a rapid reorganisation of the gene pool when a population passes through a bottleneck in population size, and is usually accompanied by rapid and drastic morphological changes (Mayr, 1981).

Having gained by chance an enhanced ability and propensity for vocalisation, their infants were exposed to a large range of sounds which they would eventually imitate. The natural ability of the developing brain to categorise and recognise features would result in the perception of these sounds as discrete units. This is what one would expect from an adaptive neural net.

Having developed a group phonology, sounds and clusters of sounds would become arbitrarily associated with meanings as individuals engaged in basic gestural communication. This enhanced communication would have conferred survival benefits and encouraged social interaction. It is also likely that the more expressive individuals were more successful reproducers, the "smooth talker" effect.

At this stage the proto-language could be seen as an information virus inhabiting, in an opportunistic way, those parts of the brain close to the sites involved in hearing and articulation. The synergistic relationship between language, biology, and society would have ensured a rapid evolution of each to the benefit of the others. Language optimised human biology for speech, while the existing organisation of the brain optimised language itself. Social encouragement and survival pressures reinforced both language and the anatomy of articulation. This model implies that language and human biology evolved in step and very rapidly.

To appreciate the possible rapidity of the structuring and evolution of the proto-language into a simple but fully grammatical language, consider that an unstructured pidgin, used by diverse linguistic groups to communicate, can evolve into a creole with a fully developed and original grammar within two generations (Kess, 1990). This structure arises from the way in which the brain organises linguistic information. It is not inconceivable that the evolution of a simple but generative language, along with the basic features of modern anatomy, occurred within, say, one hundred generations of isolation. Such an event would appear punctuational on the Pleistocene time scale.

Language would have evolved in a rapid sequence of revolutions. A chance modification of anatomy gave rise to a phonology. As soon as concatenations of sounds became long enough and rapid enough they became percieved as units and started to be associated with meanings to become words. The number of words increased gradually, perhaps quite rapidly once the advantages were understood, until the number of words and the speed of concatenation were sufficient to produce grammar. The transitions from phonemes to words, and from words to sentences, would each have occurred within two generations, while the periods in between may have been several generations. The transitions would have occured as a result of the general ability of the brain to recognise patterns and categorise information. The periods between transitions would have been characterised by steady, perhaps conscious, improvement in speed, accuracy, and number of spoken units, along with first acoustic then semantic feature mapping and categorisation laying the foundation for transitions to higher levels of organisation.

On the face of it, anatomically modern humans were less robust than their ancestors or competitors, with fragile crania, a much greater risk of choking due to the low position of the larynx, difficulties giving birth to the large cranium, and problems with compacted wisdom teeth due to shortening of the jaw. There must have been a powerful advantage to allow these relatively weak creatures to survive while their more robust competitors failed. The sudden appearance of anatomical adaptations most of which could only be useful for speech, and which were a disadvantage in many other ways, argues powerfully for the development of generative spoken language at this time.


Language is a very complex phenomenon and it is understandable to assume that it evolved over a very long period. Such an assumption is most often made, but is not necessarily true. This paper views language as a product of the emergent properties of very large adaptive neural nets. It was bound to evolve suddenly and rapidly as soon as the chance deviation of the vocal tract allowed for sufficient speed and clarity of articulation.

Language evolved as a social phenomenon, increasing the fitness of the individual only within the social context, and increasing the fitness of the society as a whole. Language conferred such a social advantage that it could evolve at the expense of the fitness of the isolated individual.

The structure of language reflects the ordering principles of the brain. It is not necessary to postulate a radical new mode or process of thought, the ATO (Klein, 1989), or GAD (Corballis, 1991). Although language represents a great enhancement of thought, it is processed by pre-existing neural circuitry in a manner which is not fundamentally different to the type of spatial and temporal processing which would have been used in pre-language thought.

This paper proposes that anatomically modern humans emerged suddenly with a fully generative language after a relatively short period of isolation and rapid evolution. Also that it is unlikely that any other species of Homo had evolved any form of generative communication system.

Supporting evidence is likely to come from the analysis of genetic and linguistic trees.


Dr Colin Groves, for kind advice and encouragement.


Allinson N and Johnson M (1989). Realisation of self organising neural maps in {0,1}n space. In: Taylor J and Mannion C (eds). New Developments in Neural Computing. New York, Adam Hilger, pp 79-85

Cavalli Sforza L (1989). The last 100,000 years of human evolution: The vantage point of genetics and archaeology. In: Hominidae: Proceedings from the 2nd International Congress of Human Paleontology. Milan, Jaca Book, pp 409-412.

Corballis M (1991).The Lopsided Ape. New York, Oxford University Press, pp 280-304.

Davidson I and Noble W (1989).The archaeology of perception: Traces of depiction and language. Current Anthropology

Duchin L (1990). The evolution of articulate speech: Comparative anatomy of the oral cavity in Pan and Homo. Journal of Human Evolution 19, 687-697.

Edelman G (1989). Neural Darwinism. New York, Oxford University Press, pp 143, 254.

Feldman H, Holland A, Kemp S, and Janosky J (1992). Language development after unilateral brain injury. Brain and Language 42, 89-102.

Hecht-Neilsen R (1990). Neurocomputing. New York: Addison-Wesley, pp 63-70.

Hertz J, Krogh A, and Palmer R (1990). Introduction to the Theory of Neural Computation. New York, Addison-Wesley, pp 217-237.

Jubak J (1992). In the Image of the Brain. Boston, Little, Brown & Co, pp 170-175.

Kess J (1990). Psycholinguistics. John Benjamins Publishing Company, pp 253-278.

Klein S (1989). Human cognitive changes at the middle to upper palaeolithic transition: The evidence of Boker Tachtit. In: Mellars P (ed), The Emergence of Modern Humans: An Archaeological Perspective. Edinburgh, Edinburgh University Press, pp 499-516.

Lakoff G (1987). Women, Fire, and Dangerous Things: What categories reveal about the mind. Chicago, University of Chicago Press, pp 283-303

Lieberman P, Laitman J, Reidenberg J, and Gannon P (1992). The anatomy, physiology, acoustics and perception of speech: Essential elements in analysis of the evolution of human speech. Journal of Human Evolution 23, 447-467.

Mayr E (1982). Processes of speciation in animals. In: Barigozzi C (ed), Mechanisms of Speciation. New York, Alan R Liss, Inc, pp 1-19.

Noble W and Davidson I (1993). Tracing the emergence of modern human behaviour: Methodological pitfalls and a theoretical path. Journal of Anthropological Archaeology 12, 121-149

Rumelhart D (1989). Brain-style computation: Mental processes emerge from interactions among neuron-like elements. In Brink J and Haden C (eds). The Computer and the Brain: Perspectives on human and artificial intelligence. North-Holland, Elsevier Science Publishers, pp 111-121.

Stringer C, Hublin J, and Vandermeersch B (1984). The origins of anatomically modern humans in Western Europe. In: The Origins of Modern Humans: A world survey of the fossil evidence. New York, Alan R Liss Inc, pp 51-135.

Swisher C, Curtis G, Jacob T, Getty G, Suprijo A, and Widiasmoro (1994). Age of the earliest known hominids in Java, Indonesia. Science 263, 1118-1121

First published by the Australasian Society for Human Biology.

Other Papers

Page Authored by MacKaos Consulting.