Language Learning

Language Learning


Algorithmic Information Theory (AIT)

A key idea in AIT is that a body of information is not random if it can be generated by a computer program which is smaller than that body of information.

This elegant idea makes a connection between computing and information compression and this suggests, on a superficial reading, that there is no difference between AIT and the SP theory.

But workers in AIT have not suggested that 'computing' in itself might be understood as information compression. There is no suggestion that the concept of an 'algorithm' might be understood in those terms (see, for example, Uspensky, V. A., J Symbolic Logic 57(2), 385-412, 1992).

It is not very surprising, therefore, that AIT does not provide any resolution of the apparent conflict between 'computing as compression' and the obvious fact that computers can be used to de-compress information and create redundancy.

The way in which 'computing as compression' may be reconciled with the fact that computers can be used to create redundancy is discussed in Computing and information compression: a reply and Computing as compression: an overview of the SP theory and system with other observations on the relationship between the SP theory and AIT. An example of how compression can achieve the effect of generation is given in Parsing as information compression by multiple alignment, unification and search: SP52.

Minimum Length Encoding (MLE)

Information compression has been a significant theme in research on inductive inference including grammatical inference under the banner of Minimum Length Encoding (MLE) pioneered by Solomonoff, Wallace and Boulton, Rissanen and others. MLE is an umbrella term covering the closely-related concepts of Minimum Message Length encoding (MML) and Minimum Description Length encoding (MDL). 

The key idea in MLE is that in grammatical inference and related processing, one should seek to minimise (G + E), where G is the size of the grammar (in bits or comparable measure) and E is the size of the raw data from which the grammar is derived, after the raw data has been encoded in terms of the grammar.

This idea seems to produce a happy compromise between grammars that are, at one extreme, very small but are inefficient for encoding data and, at the other extreme, grammars that are very efficient for encoding data but are excessively large.

This idea provides an explanation of how it is that children can make grammatical generalisations and eventually correct their erroneous 'over' generalisations (e.g. "He hitted me", "Look at those mouses") without apparently needing explicit correction from other people. Both kinds of generalisation, by definition, have zero frequency in the child's experience but MLE provides a means of distinguishing 'good' ones' from 'bad' ones without external supervision of the learning process (see Language Learning as Compression).

Distinctive features of the SP programme compared with AIT and MLE

The SP programme of research is based on MLE principles and is related to concepts in AIT but it has distinctive features compared with these fields:
  • The SP programme seeks to integrate all kinds of computing and formal reasoning within a framework of information compression. The goal is broader than it is in other research in AIT or MLE.
  • The SP programme is based on the hypothesis that all kinds of compression may be understood in terms of information compression by multiple alignment, unification and search (ICMAUS). In essence this means that all kinds of information compression is achieved by the unification of matching patterns. All existing and projected SP models are restricted to ICMAUS mechanisms and avoid 'arithmetic coding' and related techniques which are used in some research on MLE.
    The restriction has been imposed in the interests of simplicity in the SP theory. If, as conjectured, arithmetic may be understood in terms of ICMAUS then arithmetic coding may also be understood in those terms. But until this has been demonstrated, the adoption of arithmetic coding or any similar technique would add unwanted complexity to the SP model.
  • In the first point above, the phrase "all kinds of computing" includes the concept of 'computing' itself in its full generality. Thus the SP programme hypothesises that other models of computing such the Turing model or the Post Canonical System may be understood within the ICMAUS framework. By contrast, research in AIT and MLE accepts the Turing model (and equivalent models) as the basis of all kinds of computing.



Language Learning