These papers try to understand the computations which go on in the brain, by building computational models and comparing them with observation. All can be viewed here as web pages; most of them are published, and you can get copies of any paper from me on request (paper, rtf or postscript). If you read any of them in detail, please take a few moments to email your comments to me.
Keywords: Cognition, representation, evolution, computational modelling, optimality, social intelligence, navigation, language, learning, consciousness.
Key results of each paper :
|
A Theory of Learning for Categorial Grammars: Categorial grammars are a unification-based grammar formalism which describe many complex features of adult language simply and neatly. Each word is represented by a re-entrant feature structure, and the whole of a language is defined by its word feature structures. This paper presents a simple working theory of how word feature structures are learnt by children from sentences they hear. The sound, meaning and syntax of any word can be learnt robustly from about six examples of its use. Language regularities and exceptions are learnt by the same mechanism. |
| . | |
|
Words, Memes and Language Evolution: In the theory of language learning described above, each word is a parasitic species (or meme) which propagates from one brain to another by the learning mechanism. Over many generations, word species evolve, leading to language change. Selection pressures on words include useful meaning, productivity, ambiguity, economy, ease of learning, and social identification. Word evolution accounts for many prominent features of individual languages, and for language universals such as some Greenberg universals and the Chomskyan 'Head parameter'. So these features of language tell us about word evolution, not about innate structures in the human brain. They tell us less than we thought about the language module in the mind, but less is more. |
| . | |
|
A Theory of Language Learning The theory described in 'A Theory of Learning for Categorial Grammars' is expanded and compared with known facts of childrens' language learning. It agrees with many observations - such as the rapidity and robustness of learning, the order of learning different parts of speech, early verb-centred syntax, productive generalisations, structure-dependent learning, transient errors such as over-regularisation, cue strength effects, auxiliaries and complement-taking verbs. |
| . | |
|
Hybrid Cognition It is widely assumed that all computation in the brain is done by neurons, but we should periodically question this assumption. Possibly the strongest selection pressure on the brain arises from the need to plan movements, making them appropriate for the shapes and positions of objects around us. This requires an unrestricted 3-D metric representation of local space. I argue that neurons alone cannot do this well, but a hybrid mechanism, using wave-like storage of information in the thalamus, does so. If the wave storage is the basis of phenomenal consciousness, this gives a highly constrained and predictive theory of consciousness - one which agrees well with the data. |
| . | |
|
The Evolution of Language from Social Intelligence : Proposes that language evolved from, and is an application of, primate social intelligence. This agrees with the evolutionary constraint (that language cannot have evolved from scratch in homo sapiens; see 'a speed limit for evolution' below), with many facts about the structure and use of language (e.g. language requires a theory of mind), and with neurophysiology. The computational model of primate social intelligence (above) extends to be a working model of language learning and use. In this model, language meanings are scripts, and words are script functions. |
| . | |
|
Primate Social Intelligence Primates have much greater social intelligence than any other land mammal; they know each other as individuals, know about kinship relations, rank, alliances, false alarm calls and a lot more. This paper presents a computational model of that social intelligence, which uses tree-structured scripts to represent social situations. The model explains how primates learn and use regularities of social life. It is compared with observations of primates, particularly from Cheney & Seyfarth's How Monkeys See the World. |
| . | |
|
A Speed Limit for Evolution : Evolution creates design information,but does so slowly. This paper proves a speed limit on the rate at which genetic information can accumulate by natural selection - typically the rate must be less than one bit per generation. This implies that the differences in design between the human brain and the chimpanzee brain amount to less than 5 KBytes of useful design information, or a small fraction of the total; which is not enough to have evolved a language acquisition device (LAD) ab initio in mankind. |
| . | |
|
An Optimal Yardstick for Cognition Brains cannot carry on getting better and better without limit; there is a best possible brain for any species. This paper derives a mathematical form for the input/output relation of the best possible brain. Most species are near to this optimum. This fact leads to useful predictions for many domains of cognition - eg that brains have internal representations of external reality, and use a Bayesian criterion to learn regularities as fast as possible. |
| . | |
|
Navigation by Fragment Fitting : Discusses how animals navigate, fitting together map 'fragments' like pieces of a jig-saw, to form a mental map of their environment. Describes a working computational model of fragment fitting, and comparisons with data on hippocampal place cells. |
| . |
Please Email comments or questions to Robert Worden.
See a short biography.
Charteris Ltd, 6 Kinghorn Street, London EC1A 7HT
Updated 15 October 1997