Machine Learning and the Acquisition of Natural Concepts based on the Psychological Theory of Prototypes

This thesis describes a research programme on machine learning, based on the fundamental process of categorisation. This work draws upon the cognitive psychological theory of prototypical concepts. This theory states that concepts, learnt naturally through interaction with the environment (basic concepts or categories), are not structured or defined in logical terms but are clustered in accordance with their similarity to a central prototype, representing the "most typical" member. The PLANC (Prototypical Learner & Acquirer of Natural Concepts) system, built to operationalise aspects of Prototype theory, differs from mainstream Machine Learning in three respects. First, examples are provided visually in the form of two-dimensional canonical views of three-dimensional objects, drawn from a variety of generic concepts. The number and characteristics of these concepts being unknown to PLANC. Secondly, no form of instruction is given as to what the target concepts should be, these being detected by using only the inherent correlation of perceptual attributes. Thirdly, the final concept is represented as a non-logical description consisting of a structural description abstracted from category members, together with typical shapes of individual parts. The PLANC system incorporates ideas and techniques from a variety of other fields: Machine Learning, Pattern Recognition, Image Processing, Computer Vision, Statistics and Cognitive Psychology. These are embodied in a set of modules that perform the tasks of: decomposing shapes into component parts, describing the structure between parts, and categorisation based upon similarities of structure and shape of parts. Results produced by PLANC indicate that concepts held by humans arise automatically from the detection of similarities between the innate structure and shape of natural objects. This provides strong support for Prototype theory as it shows the depth of structuring of the real world that must exist for natural concepts to be recoverable from such an impoverished source of information as two-dimensional silhouettes.