A Runtime Switchable Multi-Phase Convolutional Neural Network for Resource-Constrained Systems

Convolutional Neural Networks (CNNs) are widely used in various systems, including resource-constrained embedded systems or IoT devices. In such it is typical to deploy compressed pruned CNNs, instead of original ones, at the cost reduced accuracy. Existing CNN pruning techniques have primarily focused on minimizing resource requirements. However, today’s increasingly dynamic both demands and availability. Thus, previous that only consider given static cases no longer efficient. this paper, we propose a novel multi-phase enables multi-objective exploration number candidates out single CNN. proposed technique, can operate versions depending which subsets weights be transformed one best matches constraint adaptively efficiently. For that, first sparsest form; then set parameters (sub-network) additionally supplemented as phase goes by. As result, network for all different phases represented by they form pareto solution over accuracy usage trade-off. work, target CPU-based inference engines most do not luxury specialized co-processor support GPUs HW accelerators. The technique has been implemented publicly available CPU engine, Darknet, its effectiveness validated with popular terms design space capability runtime switchability.