2021/11/13 by Priyesh Shukla, Shamma Nasrin, Shukla, Priyesh +7
Computer Science · Engineering · #Advanced Memory and Neural Computing #Advanced Neural Network Applications #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #FOS: Electrical engineering #Hardware Architecture (cs.AR) #Image and Video Processing (eess.IV) #Machine Learning (cs.LG) #Robotics (cs.RO) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2111.07125
openalex publication_date 2021/11/13 · openalex created_date 2022/10/20 · openalex updated_date 2026/07/28
We propose MC-CIM, a compute-in-memory (CIM) framework for robust, yet low\npower, Bayesian edge intelligence. Deep neural networks (DNN) with\ndeterministic weights cannot express their prediction uncertainties, thereby\npose critical risks for applications where the consequences of mispredictions\nare fatal such as surgical robotics. To address this limitation, Bayesian\ninference of a DNN has gained attention. Using Bayesian inference, not only the\nprediction itself, but the prediction confidence can also be extracted for\nplanning risk-aware actions. However, Bayesian inference of a DNN is\ncomputationally expensive, ill-suited for real-time and/or edge deployment. An\napproximation to Bayesian DNN using Monte Carlo Dropout (MC-Dropout) has shown\nhigh robustness along with low computational complexity. Enhancing the\ncomputational efficiency of the method, we discuss a novel CIM module that can\nperform in-memory probabilistic dropout in addition to in-memory weight-input\nscalar product to support the method. We also propose a compute-reuse\nreformulation of MC-Dropout where each successive instance can utilize the\nproduct-sum computations from the previous iteration. Even more, we discuss how\nthe random instances can be optimally ordered to minimize the overall\nMC-Dropout workload by exploiting combinatorial optimization methods.\nApplication of the proposed CIM-based MC-Dropout execution is discussed for\nMNIST character recognition and visual odometry (VO) of autonomous drones. The\nframework reliably gives prediction confidence amidst non-idealities imposed by\nMC-CIM to a good extent. Proposed MC-CIM with 16x31 SRAM array, 0.85 V supply,\n16nm low-standby power (LSTP) technology consumes 27.8 pJ for 30 MC-Dropout\ninstances of probabilistic inference in its most optimal computing and\nperipheral configuration, saving 43% energy compared to typical execution.\n