In this paper, we present a new approach to learning cascaded classifiers foruse in computing environments that involve networks of heterogeneous andresource-constrained, low-power embedded compute and sensing nodes. We presenta generalization of the classical linear detection cascade to the case oftree-structured cascades where different branches of the tree execute ondifferent physical compute nodes in the network. Different nodes have access todifferent features, as well as access to potentially different computation andenergy resources. We concentrate on the problem of jointly learning theparameters for all of the classifiers in the cascade given a fixed cascadearchitecture and a known set of costs required to carry out the computation ateach node.To accomplish the objective of joint learning of all detectors, wepropose a novel approach to combining classifier outputs during training thatbetter matches the hard cascade setting in which the learned system will bedeployed. This work is motivated by research in the area of mobile health whereenergy efficient real time detectors integrating information from multiplewireless on-body sensors and a smart phone are needed for real-time monitoringand delivering just- in-time adaptive interventions. We apply our framework totwo activity recognition datasets as well as the problem of cigarette smokingdetection from a combination of wrist-worn actigraphy data and respirationchest band data.
展开▼
