The networks considered in this paper consist of tiny energy constrained commodity sensors massively deployed, along with one or more sink nodes providing interface to the outside world. Our contribution is to propose a scalable energy-efficient training protocol for nodes that are initially anonymous, asynchronous and unaware of their locations. Training partitions the nodes into clusters where data can be gathered from the environment and synthesized under local control. Further this training provides a virtual tree for efficient communication routing from clusters to the sink. Being energy-efficient, our training protocol can be run on either a scheduled or ad-hoc basis to provide robustness and dynamic reconfiguration.