Search

The use of hardware accelerators, based on code and data offloading devoted to overcoming the CPU limitations in cores, is one of the main distinctive trends in high-end computing and related applications in the last decade. However, while code offloading is convenient for performance improvement, becoming a commonly used paradigm, memory access and management are a source of bottlenecks due to the need to interact with different address spaces. In this regard, NVidia introduced the CUDA Unified Memory model to avoid explicit memory copies between the machine hosting the accelerator device and the device itself and vice-versa. This paper shows a novel design and implementation of the support to the CUDA Unified Memory in open-source GPGPU virtualization services. The performance evaluation demonstrates that the overhead due to the virtualization and remoting is acceptable considering the possibility of sharing CUDA-enabled GPUs between various and heterogeneous machines hosted at the edge, in cloud infrastructures, or as accelerator nodes in an HPC scenario. A prototype implementation of the proposed solution is available as open-source.

Retinal prosthetic systems have been developed to help blind patients suffering from retinal degenerative diseases gain some useful form of vision. Various experimental and computational studies have been performed to test electrical stimulation strategies that can improve the performance of these devices. Detailed computational models of retinal neurons, such as retinal ganglion cells (RGCs) and bipolar cells (BCs), allow us to explore the mechanisms underlying the response of cells to electrical stimulation. While electrophysiological studies have shown the presence of voltage-gated ionic channels in different regions of BCs, many of the existing cone BCs models are assumed to be passive or only contain calcium channels at the synaptic terminals. We have utilized our Admittance Method (AM)-NEURON computational platform to implement a more realistic model of ON-BCs. Our model closely replicates the recent patch-clamp experiments directly measuring the response of ON-BCs to epiretinal electrical stimulation and thereby predicts the regional distributions of the ionic channels. Our computational results further indicate that outward potassium current strongly contributes to the depolarizing voltage transient of ON-BCs in response to electrical stimulation.

The LoRa modulation scheme is becoming one of the most adopted Internet of Things wireless physical layer due to its ability to transmit data over long distances with low power requirements. Typical demodulation techniques for LoRa utilize variants of non-coherent Frequency Shift Keying demodulation. This paper aims to capitalize on the robustness of deep learning techniques, specifically by using convolutional neural networks to demodulate LoRa symbols. We achieve this by building a dataset consisting of emulated time-domain LoRa symbols across a range of channel impairments; namely, we examine additive white Gaussian noise together with carrier frequency offset and time offset. The presented results show an improvement when utilizing deep learning over typical non-coherent detection while performing very close to the optimal matched filter.

Spiking neural networks (SNNs) operating with sparse binary signals (spikes) implemented on event-driven hardware can potentially be more energy -efficient than traditional artificial neural networks (ANNs). However, SNNs perform computations over time, and the neuron activation function does not have a well-defined derivative leading to unique training challenges. In this paper, we discuss the various spike representations and training mechanisms for deep SNN s. Additionally, we review applications that go beyond classification, like gesture recognition, motion estimation, and sequential learning. The unique features of SNNs, such as high activation sparsity and spike-based computations, can be leveraged in hardware implementations for energy-efficient processing. To that effect, we discuss various SNN implementations, both using digital ASICs as well as analog in-memory computing primitives. Finally, we present an outlook on future applications and open research areas for both SNN algorithms and hardware implementations.

The use of robotics in health care has seen a recent rise in interest due to its potential for use during the SARS-CoV-2 pandemic. The transmission rate of COVID-19 has meant that health-care workers are under increasing pressure, risks, and workload to manage the requirements of personal protective equipment, strict disinfection procedures, and the heightened medical needs of patients. Patients are suffering from isolation, and not just in hospitals: higher-risk individuals must shelter, meaning social interactions, particularly in care homes, are limited. Robots can help by providing disinfection and logistics services that support patients and health-care professionals, by acting as devices to be used for rehabilitation at home (for both pre-existing conditions and for COVID-19-related treatment), and via interventional systems that can widely distribute future vaccinations.

This paper presents the idea of using LoRaWAN based sensors in isolated areas with intermittent gateway presence. The intent is to expand the network coverage to areas with no preexisting infrastructure, and automate the data collection process in Delay Tolerant Network (DTN) scenarios. The use case considered in the paper is that of ocean monitoring, where sensor nodes can be deployed far from the shore to record ocean-related data, while gateways located on ships collect the sensors’ telemetry when the ship passes close by the nodes. We discuss the LoRaWAN capabilities and limitations for such scenario, focusing in particular on the energy consumption on the end nodes’ side. Finally, we present an optimized strategy that optimizes the duty cycle for these remote nodes. The experiments show that by adopting this solution, the nodes spend less energy, which leads to their lifespan to be extended significantly.

Watermarking is a process in which both physical and digital media are marked using watermarks in order to protect ownership of the watermarked media. Digital watermarking is a technique where a watermark gets embedded into the carrier signal while preserving the quality of the original media. Embedding can happen in various domains and could be both hidden and plain, but the quality and the information carried by the signal should not deteriorate. This paper deals with hiding watermarks into speech audio signals using deep neural networks. We present an encoder-decoder architecture that achieved PSNR value greater than 57dB, which we used as a preservation measure of the original signal and message transmission accuracy of almost 100%. Audio data, used in this paper, consists of speeches from the Parliament of Montenegro.

One of the main goals of knowledge discovery from environmental data is through data analysis to find the relationship between the living organisms, represented with the diversity of the diatoms community members, and the characteristics of the environment. This is very important information for both ecologists and decision makers. Therefore, in this paper we apply various machine learning algorithms for revealing this relationship by using different number of discretization levels for the target attribute. The target attribute represents the biodiversity index of the community and it is calculated based on the abundances of the diatoms. For building models, different types of machine learning algorithms are considered including decision trees, rule induction algorithms, neural networks and Naive Bayes. The obtained models are also examined regarding resistance to over-fitting, as well as statistical significance.

The effects of air pollution on people, the environment, and the global economy are profound - and often under-recognized. Air pollution is becoming a global problem. Urban areas have dense populations and a high concentration of emission sources: vehicles, buildings, industrial activity, waste, and wastewater. Tackling air pollution is an immediate problem in developing countries, such as North Macedonia, especially in larger urban areas. This paper exploits Recurrent Neural Network (RNN) models with Long Short-Term Memory units to predict the level of PM10 particles in the near future (+3 hours), measured with sensors deployed in different locations in the city of Skopje. Historical air quality measurements data were used to train the models. In order to capture the relation of air pollution and seasonal changes in meteorological conditions, we introduced temperature and humidity data to improve the performance. The accuracy of the models is compared to PM10 concentration forecast using an Autoregressive Integrated Moving Average (ARIMA) model. The obtained results show that specific deep learning models consistently outperform the ARIMA model, particularly when combining meteorological and air pollution historical data. The benefit of the proposed models for reliable predictions of only 0.01 MSE could facilitate preemptive actions to reduce air pollution, such as temporarily shutting main polluters, or issuing warnings so the citizens can go to a safer environment and minimize exposure.

Mobile Edge Computing (MEC) is a network architecture that takes advantage of resources available at the edge of the network to enhance the mobile user experience by decreasing the service latency. MEC solutions need to dynamically allocate the requests as close as possible to their users to avoid high latency. However, the request allocation does not depend only on the geographical location of the servers, but also on their requirements. The task of choosing and allocating appropriate servers in a MEC environment is challenging because it involves many parameters. This paper proposes a Stochastic Petri Net (SPN) model to represent a MEC scenario and analyze its performance. The model focuses on parameters that can directly impact the service Mean Response Time (MRT) and resource utilization level. We propose case studies with numerical analyzes using real-world values to validate the proposed model. The main objective is to provide a practical guide to assist infrastructure administrators to adapt their architectures, finding a trade-off between MRT and resource utilization level.

In this work, we investigate information freshness in a status update communication system consisting of a source-destination link. Initially, we study the properties of a sample path of the age of information (AoI) process at the destination. We obtain a general formula of the stationary distribution of the AoI, under the assumption of ergodicity. We relate this result to a discrete time queueing system and provide a general expression of the generating function of AoI in relation with the system time and the peak age of information (PAoI) metric. Furthermore, we consider three different single-server system models and we obtain closed-form expressions of the generating functions and the stationary distributions of the AoI and the PAoI. The first model is a first-come-first-served (FCFS) queue, the second model is a preemptive last-come-first-served (LCFS) queue, and the last model is a bufferless system with packet dropping. We build upon these results to provide a methodology for analyzing general non-linear age functions for this type of systems, using representations of functions as power series.