Beyond energy, the growing number of defects in physical substrates is becoming another major constraint that affects the design of computing devices and systems. As the underlying semiconductor technologies are getting less and less reliable, the probability that some components of computing devices fail also increases, preventing designers from realizing the full potential benefits of on-chip exascale integration derived from near atomic scale feature dimensions. As the quest for performance confronts permanent and transient faults, device variation, and thermal issues, major breakthroughs in computing efficiency are expected to benefit from unconventional and new models of computation, such as brain-inspired computing. The challenge is then to find not only high-performance and energy-efficient, but also fault-tolerant computing solutions. Neural computing principles remain elusive, yet as source of a promising fault-tolerant computing paradigm. In the quest to fault tolerance can be translated into scalable and reliable computing systems, hardware design itself and/or to use circuits even with faults has further motivated research on neural networks, which are potentially capable of absorbing some degrees of vulnerability based on their natural properties. This paper presents a survey on fault tolerance in neural networks manly focusing on well-established passive techniques to exploit and improve, by design, such potential but limited intrinsic property in neural models, particularly for feedforward neural networks. First, fundamental concepts and background on fault tolerance are introduced. Then, we review fault types, models, and measures used to evaluate performance and provide a taxonomy of the main techniques to enhance the intrinsic properties of some neural models, based on the principles and mechanisms that they exploit to achieve fault tolerance passively. For completeness, we briefly review some representative works on active fault tolerance in neural networks. We present some key challenges that remain to be overcome and conclude with an outlook for this field.

Convolutional neural networks (CNN) have turned into one of the key algorithms in machine learning for content classification of digital images. Nevertheless, the CNN computational complexity is considerable larger than classic algorithms, thus, CPU- or GPU-based platforms are generally used for CNN implementations in many applications, but often do not fulfill portable requirements due to resources, energy and real-time constrains. Therefore, there is a growing interest on real time processing solutions for object recognition using CNNs mainly implemented on embedded systems, which are limited both in resources and energy consumption. An updated review of prominent reported approaches for mapping CNNs onto embedded systems is described in this paper. Two main solutions trends for reducing the hardware CNN workload are distinguished through a deduced taxonomy. One is focused on algorithm level solutions to reduce the number of multiplications and CNN coefficients. On the other hand, hardware level solutions goal is to achieve processing time, power consumption and hardware resources reduction. Two dominant hardware level design strategies are pointed out as oriented to either reducing the energy consumption and resources utilization meeting real-time requirements or increasing the throughput at the expense of resources utilization. Finally, two identified design strategies for CNN hardware accelerators are proposed as opportunity research areas.

Mobile Edge Computing (MEC) relates to the deployment of decision-making processes at the network edge or mobile devices rather than in a centralized network entity like the cloud. This paradigm shift is acknowledged as one key pillar to enable autonomous operation and self-awareness in mobile devices in IoT. Under this paradigm, we focus on mobility-based services (MBSs), where mobile devices are expected to perform energy-efficient GPS data acquisition while also providing location accuracy. We rely on a fully on-device Cognitive Dynamic Systems (CDS) platform to propose and evaluate a cognitive controller aimed at both tackling the presence of uncertainties and exploiting the mobility information learned by such CDS toward energy-efficient and accurate location tracking via mobility-aware sampling policies. We performed a set of experiments and validated that the proposed control strategy outperformed similar approaches in terms of energy savings and spatio-temporal accuracy in LBS and MBS for smartphone devices.

In this paper we investigate the impact of data uncertainty on the prediction of vital signs abnormalities. We analyze the evolution of mean arterial blood pressure (MAP) and heart rate (HR) vital signs time series over a given forecasting window, whose values are classified according to a model of normality that models patient’s physiological deterioration. To this end, we developed a generic prediction framework that enables the deployment of different methods for data preparation, training, prediction, and classification so as to perform a set of experiments. To address our purpose, we selected a general regression neural network model (to predict vital sign trends) and a support vector machine (to recognize abnormalities) as the appropriate methods for a baseline analysis. However, other algorithm choices could be tailored in the defined prediction framework. We extract from the Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC II) waveform database a set of 302,940 vital sign samples (organized as 459 records) to define training and testing datasets. From our experiments, we obtained evidence demonstrating the incidence of null values in the MAP/HR input time series on forecasting tasks. This was performed under different conditions in terms of null values in the observation window, and varying other factors like labeling methods and forecast window sizes. In our experiments, we concentrate on analyzing the extent of prediction performance under such conditions, and identify that a tolerable level of null values is around 25% where it is possible to still derive forecast windows with an accuracy target ≥0.9 for both MAP and HR vital signs. The presented evaluation contributes to acknowledge the importance of uncertainty conditions on the prediction performance of vital signs, and paves the way for future extensions involving other prediction approaches.

Convolutional Neuronal Networks (CNN) implementation on embedded devices is restricted due to the number of layers of some CNN models. In this context, this paper describes a novel architecture based on Layer Operation Chaining (LOC) which uses fewer convolvers than convolution layers. A reutilization of hardware convolvers is promoted through kernel decomposition. Thus, an architectural design with reduced resources utilization is achieved, suitable to be implemented on low-end devices as a solution for portable classification applications. Experimental results show that the proposed design has a competitive processing time and overcomes resource utilization when compared with state-of-the-art related works.