Saligrama, Venkatesh2019-08-302019-08-302019-04-01Venkatesh Saligrama. 2019. "Cost aware Inference for IoT Devices." Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS) 2019, Naha, Okinawa, Japan. PMLR: Volume 89.https://hdl.handle.net/2144/37592Networked embedded devices (IoTs) of limitedCPU, memory and power resources are revo-lutionizing data gathering, remote monitoringand planning in many consumer and businessapplications. Nevertheless, resource limita-tions place a significant burden on their ser-vice life and operation, warranting cost-awaremethods that are capable of distributivelyscreening redundancies in device informationand transmitting informative data. We pro-pose to train a decentralized gated networkthat, given an observed instance at test-time,allows for activation of select devices to trans-mit information to a central node, which thenperforms inference. We analyze our proposedgradient descent algorithm for Gaussian fea-tures and establish convergence guaranteesunder good initialization. We conduct exper-iments on a number of real-world datasetsarising in IoT applications and show that ourmodel results in over 1.5X service life withnegligible accuracy degradation relative to aperformance achievable by a neural network.Copyright 2019 by the author(s).Conference materials0000-0002-0675-2268 (Saligrama, Venkatesh)466437