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