ENOS: Energy-Aware Network Operator Search in Deep Neural Networks

This work proposes a novel Energy-aware Network Operator Search (ENOS) approach to address the energy-accuracy trade-offs of deep neural network (DNN) accelerator. In recent years, hardware-friendly inference operators such as binary-weight, multiplication-free, and deep-shift have been proposed improve computational efficiency DNN However, simplifying invariably comes at lower accuracy, especially on complex processing tasks. While prior works generally implement same operator throughout architecture, ENOS framework allows an optimal layer-wise integration with optimal precision maintain high prediction accuracy energy efficiency. The search in is formulated continuous optimization problem, solvable using gradient descent methods, thereby minimally increasing training cost when learning both weights. Utilizing ENOS, we discuss multiply-accumulate (MAC) cores for digital spatial architectures that can be reconfigured different varying computing precision. methods single bi-level objectives are discussed compared. We also sequential assignment strategy only learns one layer step. Furthermore, stochastic mode presented. characterized ShuffleNet SqueezeNet CIFAR10 CIFAR100. Compared conventional uni-operator approaches, under budget, improves by 10–20%. outperforms comparable mixed-precision implementations 3-5% budget.