AISB50: Artificial Intelligence at a new branch point

Tools

Bishop, Mark (J. M.). 2014. AISB50: Artificial Intelligence at a new branch point. The AISB Quarterly, 139, pp. 11-17. ISSN 0268-4179 [Article]

Preview

Text
On AISB50 Short for Q.pdf - Accepted Version
Available under License Creative Commons Attribution Non-commercial.
Download (104kB) | Preview

Official URL: http://aisb.org.uk/publications/aisbq/AISBQ139.pdf

Abstract or Description

It should be noted that from now on ‘the system’ means not the nervous system but the whole complex of the organism and the environment. Thus, if it should be shown that ‘the system’ has some property, it must not be assumed that this property is attributed to the nervous system: it belongs to the whole; and detailed examination may be necessary to ascertain the contributions of the separate parts. W. Ross Ashby, 1952 [1] An oft repeated aphorism is that the universe is constantly changing and hence that our world is in a perpetual state of flux. In order to behave intelligently within this varying natural environment, any system be it man, machine or animal faces the problem of perceiving invariant aspects of a world in which no two situations are ever exactly the same. Cartesian theories of perception can be broken down into what Chalmers [5] calls the ‘easy problem’ of perception the classification and identification of sense stimuli and a corresponding ‘hard problem’ the realisation of the associated phenomenal state. The difference between the ‘easy’ and the ‘hard’ problems and an apparent lack of link between theories of the former and an account of the latter has been termed the ‘explanatory gap’ [10] and this [unbridgeable] gap is symptomatic of the underlying dualism. Many current theories of natural visual processes are grounded upon the idea that when we perceive, sense data is processed by the brain to form an internal representation of the world. The act of perception thus involves the activation of an appropriate representation. The easy problem reduces to forming a correct internal representation of the world and the hard problem reduces to answering how the activation of a representation gives rise to a sensory experience. In machine perception progress in solving even the ‘easy’ problem has so far been unexpectedly slow; typical bottom-up (or data driven) methodologies involve the processing of raw sense data to extract a set of features; the binding of these features into groups then classifying each group by reference to a putative set of models. Conversely, in top down methods, a typical set of hypotheses of likely perceptions is generated; these are then compared to a set of features in a search for evidence to support each hypothesis.

Item Type:

Article

Keywords:

Theory, Artificial intelligence, Machine perception, Artificial Intelligence, Nervous system structure, Typical set, Statistical classification, Top-down and bottom-up design, Face, Invariant, explanation, Classification, Perception, Realization, Face

Departments, Centres and Research Units:

Computing

Dates: