Trade-offs in the Distribution of Neural Networks in a Wireless Sensor Network

This article investigates the tradeoff between communication and memory usage in different methods of distributing neural networks in a Wireless Sensor Network. A structural approach is presented, categorized in two dimensions: horizontal and vertical decomposition. Horizontal decomposition turns out to be more attractive, due to high reuse of data present at the processor node. General properties of an alternative sematic approach are suggested theoretically allowing to dramatically increase efficiency.