Feature Detectors for Grammatical Categories

Philip Dorrell

18 February, 2006

Feature Detection: Definition

There are many parts of the brain where the activity of neurons in those parts can be understood in terms of feature detection. Feature detection is where the activity of a neuron is correlated with the detection of a particular feature.

A "feature" can be almost any aspect of the environment that is perceived. So there are visual feature detectors for colours, shapes, motion, size, number. There are aural feature detectors for pitch, loudness, rate of change of pitch and timbre of sounds.

Application to Grammar

Theories of grammar specify attributes which are used to describe both words and parts of sentences. For words we have attributes which describe grammatical category, such as noun, verb, adjective and adverb. For parts of sentences there are categories such as subject, noun phrase, verb phrase, subordinate clause.

A feature detection theory of grammar would suppose that for each of these grammatical categories there exists a set of neurons whose activity corresponds to perception of that grammatical category.

Nested Features

There is one aspect of grammar which complicates this picture, which is the recursive or nested construction of grammatical components. To give a simple example, one can have a noun phrase within a noun phrase, e.g. in:

The biggest eater of fine foods is my nephew.

"The biggest eater of fine foods" is a noun phrase which contains a sub-phrase "fine foods". If there are "noun-phrase" neurons whose activity corresponds to perception of a noun-phrase, then we might expect those neurons to respond to both phrases. But since the smaller phrase is contained within the large one, there must at a certain point be "noun-phrase" neurons responding to both phrases simultaneously.

If the grammar of human language was not recursively defined, then this problem would not arise, because no grammatical category would ever occur within itself, and neurons representing that category would never have to be active simultaneously in response to nested occurrences of that category.

How can a set of feature-detecting neurons respond to multiple occurrences of that feature? This question can be regarded as a restatement of the binding problem, which asks how the brain can bind perceptions of different types of attribute relating to one object, and not bind perceptions relating to different objects.

Conclusion

The implication is that a general solution to the binding problem was a necessary pre-condition to the development of the ability to process recursively defined syntaxes.

This paper is not a real paper – it is an example which helps to demonstrate the concept of the Web 2.0 Science Journal. (However, unlike the other example papers, it is based on a serious idea that I may write up properly sometime.) It is reviewed (in an example review) on my Web 2.0 Peer Reviews.