Successful integration of deep neural networks (DNNs) or deep learning (DL) has resulted in breakthroughs in many areas. However, deploying these highly accurate models for data-driven, learned, automatic, and practical machine learning (ML) solutions to end-user applications remains challenging. DL algorithms are often computationally expensive, power-hungry, and require large memory to process complex and iterative operations of millions of parameters. Hence, training and inference of DL models are typically performed on high-performance computing (HPC) clusters in the cloud. Data transmission to the cloud results in high latency, round-trip delay, security and privacy concerns, and the inability of real-time decisions. Thus, processing on edge devices can significantly reduce cloud transmission cost. Edge devices are end devices closest to the user, such as mobile phones, cyber–physical systems (CPSs), wearables, the Internet of Things (IoT), embedded and autonomous systems, and intelligent sensors. These devices have limited memory, computing resources, and power-handling capability. Therefore, optimization techniques at both the hardware and software levels have been developed to handle the DL deployment efficiently on the edge. Understanding the existing research, challenges, and opportunities is fundamental to leveraging the next generation of edge devices with artificial intelligence (AI) capability. Mainly, four research directions have been pursued for efficient DL inference on edge devices: 1) novel DL architecture and algorithm design; 2) optimization of existing DL methods; 3) development of algorithm–hardware codesign; and 4) efficient accelerator design for DL deployment. This article focuses on surveying each of the four research directions, providing a comprehensive review of the state-of-the-art tools and techniques for efficient edge inference.

Electroencephalogram (EEG) is a technique for recording the asynchronous activation of neuronal firing inside the brain with non-invasive scalp electrodes. Artifacts, such as eye blink activities, can corrupt these neuronal signals. While ocular artifact (OA) removal is well investigated for multiple channel EEG systems, in alignment with the recent momentum toward minimalistic EEG systems for use in natural environments, we investigate unsupervised and effective removal of OA from single-channel streaming raw EEG data. In this paper, the unsupervised wavelet transform (WT) decomposition technique was systematically evaluated for the effectiveness of OA removal for a single-channel EEG system. A set of seven raw EEG data set was analyzed. Two commonly used WT methods, Discrete Wavelet Transform (DWT) and Stationary Wavelet Transform (SWT), were applied. Four WT basis functions, namely, haar, coif3, sym3, and bior4.4, were considered for OA removal with universal threshold and statistical threshold (ST). To quantify OA removal efficacy from single-channel EEG, five performance metrics were utilized: correlation coefficients, mutual information, signal-to-artifact ratio, normalized mean square error, and time-frequency analysis. The temporal and spectral analysis shows that the optimal combination could be DWT with ST with coif3 or bior4.4 to remove OA among 16 combinations. This paper demonstrates that the WT can be an effective tool for unsupervised OA removal from single-channel EEG data for real-time applications.

Mild cognitive impairment (MCI) is the preliminary stage of dementia, which may lead to Alzheimer's disease (AD) in the elderly people. Therefore, early detection of MCI has the potential to minimize the risk of AD by ensuring the proper mental health care before it is too late. In this paper, we demonstrate a single-channel EEG-based MCI detection method, which is cost-effective and portable, and thus suitable for regular home-based patient monitoring. We collected the scalp EEG data from 23 subjects, while they were stimulated with five auditory speech signals. The cognitive state of the subjects was evaluated by the Montreal cognitive assessment test (MoCA). We extracted 590 features from the event-related potential (ERP) of the collected EEG signals, which included time and spectral domain characteristics of the response. The top 25 features, ranked by the random forest method, were used for classification models to identify subjects with MCI. Robustness of our model was tested using leave-one-out cross-validation while training the classifiers. Best results (leave-one-out cross-validation accuracy 87.9%, sensitivity 84.8%, specificity 95%, and F score 85%) were obtained using support vector machine (SVM) method with radial basis kernel (RBF) (sigma = 10/cost = 10 ² ). Similar performances were also observed with logistic regression (LR), further validating the results. Our results suggest that single-channel EEG could provide a robust biomarker for early detection of MCI.

Future Smart and Connected Communities (SCC) will utilize distributed sensors and embedded computing to seamlessly generate meaningful data that can assist individuals, communities, and society with interlocking physical, social, behavioral, economic, and infrastructural interaction. SCC will require newer technologies for seamless and unobtrusive sensing and computation in natural settings. This work presents a new technology for health monitoring with low-cost body-worn disposable fully passive electronic sensors, along with a scanner, smartphone app, and web-server for a complete smart sensor system framework. The novel wireless resistive analog passive (WRAP) sensors are printed using an inkjet printing (IJP) technique on paper with silver inks (Novacentrix Ag B40, sheet resistance of 21 mΩ/sq) and incorporate a few discrete surface mounted electronic components (overall thickness of <1 mm). These zero-power flexible sensors are powered through a wireless inductive link from a low-power scanner (500 mW during scanning burst of 100 ms) by amplitude modulation at the carrier signal of 13.56 MHz. While development of various WRAP sensors is ongoing, this paper describes development of a WRAP temperature sensor in detail as an illustration. The prototypes were functionally verified at various temperatures with energy consumption of as low as 50 mJ per scan. The data is analyzed with a smartphone app that computes severity (Events-of-Interest, or EoI) using a real-time algorithm. The severity can then be anonymously shared with a custom web-server, and visualized either in temporal or spatial domains. This research aims to reduce ER visits of patients by enabling self-monitoring, thereby improving community health for SSC.

Commercial gel ECG electrodes have a limited lifetime of operation and metal dry electrodes are rigid and noisy. In this study, we developed flexible dry ECG electrodes in different shapes and sizes using inkjet printing (IJP) technology. As an additive manufacturing technique, IJP provides the flexibility to personalize the shape and size of electrodes with rapid prototyping. IJP ECG electrodes were fabricated on flexible polyimide substrates using silver nanoparticle inks for the conductive layer and a custom polymer ink of Poly(4-vinylphenol) (PVP) for the insulator layer. The maximum and minimum diameter of the electrodes was 21 mm and 9 mm respectively. The thickness of IJP electrodes was 2 ± 0.5 μm. Various tests were performed for feasibility, including changing electrode position, and subject’s activity such as resting, standing, and walking using IJP ECG electrodes simultaneous with clinical standard gel electrodes, which provides an accurate comparison. Compared to the gel electrode, the circular shape IJP electrode was more promising than rectangular and square shapes as it showed more coherence to the gel electrode. IJP electrodes with different diameters were compared along with the gel electrode and the result demonstrated that the IJP electrode with a diameter of 21 mm, 19 mm, and 16 mm shows almost similar performance as the gel electrode. Considering ECG signal quality IJP electrode with a diameter of 19 mm was the optimal size. These IJP dry electrodes can be usable in ECG wearable as it is flexible, reusable, and can be used for a long time.