Self Organized Scale-Free Distributed Classification Using A Distributed, Unreliable Set Of Limited Capability Components

Bandwidth requirements for networks that monitor the environment often result in scale limitations. Such limitations can be avoided by applying in-network data classification to reduce required bandwidth, however, traditional strategies that require a priori network configuration become expensive as networks begin to scale. We describe a self-organizing strategy that allows a scale-free set of independent sensing components to operate as a distributed classification system. A shared storage protocol is run over a set of simple, distributed physical components, which creates an abstraction of a reliable in-network storage device. The emergent storage device is used to record classification templates. Each component that encounters environmental information uses data in the network to perform data classification, and collaborates with neighbours to increase the information content. Consensus techniques are used to disambiguate between conflicting classifications. We apply the strategy to implement a distributed face recognition implementation. The strategy allows macro-level configuration of a self organizing process that supports an application. New deployment paradigms that depend on an extremely large number of components, such as nano-technology, are well suited to benefit from the strategy.