Self-Organizing Decomposition of Functions

摘要: This paper discusses some of the issues raised by various approaches to decomposing functions and modular networks, it offers a unified framework for multiple classifier (MC) systems in general. It argues that as yet there is no general approach this problem although several provide solutions situations which parametric labelling function allows task facing classifying networks be simplified. An MC connectionist system consisting process sub-spaces within based upon similarity patterns its input domain proposed evaluated context previous broader more generally. simple automatic partitioning scheme investigated using different problems, shown effective. The degree are specialized on predictable subset overall assessed, their performance compared with equivalent single-network, undivided multiversion systems. Statistical measures ‘diversity’ previously used assess voting apply measurement specialization or bias groups sub-space nets well useful indicator across range By successively increasing overlap between partitions we show transition from experts subnets, through version sets optimal single classifiers. Finally, presented.