Next generation of embedded Information and Communication Technology (ICT) systems are interconnected and collaborative systems able to perform autonomous tasks. The remarkable expansion of the embedded ICT market, together with the rise and breakthroughs of Artificial Intelligence (AI), have put the focus on the Edge as it stands as one of the keys for the next technological revolution: the seamless integration of AI in our daily life. However, training and deployment of custom AI solutions on embedded devices require a fine-grained integration of data, algorithms, and tools to achieve high accuracy and overcome functional and non-functional requirements. Such integration requires a high level of expertise that becomes a real bottleneck for small and medium enterprises wanting to deploy AI solutions on the Edge , which, ultimately, slows down the adoption of AI on applications in our daily life. In this work, we present a modular AI pipeline as an integrating framework to bring data, algorithms, and deployment tools together. By removing the integration barriers and lowering the required expertise, we can interconnect the different stages of particular tools and provide a modular end-to-end development of AI products for embedded devices. Our AI pipeline consists of four modular main steps: (i) data ingestion, (ii) model training, (iii) deployment optimization, and (iv) the IoT hub integration. To show the effectiveness of our pipeline, we provide examples of different AI applications during each of the steps. Besides, we integrate our deployment framework, Low-Power Deep Neural Network (LPDNN), into the AI pipeline and present its lightweight architecture and deployment capabilities for embedded devices. Finally, we demonstrate the results of the AI pipeline by showing the deployment of several AI applications such as keyword spotting, image classification, and object detection on a set of well-known embedded platforms, where LPDNN consistently outperforms all other popular deployment frameworks.

We consider a group of mobile users, within proximity of each other, who are interested in watching the same online video. The common practice today is that each user downloads the video independently on her mobile device using her own cellular connection, which wastes access bandwidth and may also lead to poor video quality. We propose a novel cooperative system where each mobile device uses simultaneously two network interfaces: (i) cellular to connect to the video server and download parts of the video and (ii) WiFi to connect locally to all other devices in the group to exchange those parts. Devices cooperate to efficiently utilize all network resources and to adapt to varying wireless network conditions. In the local WiFi network, we exploit overhearing, which we further combine with network coding. The end result is savings in cellular bandwidth and improved user experience. We follow a complete approach, from theory to practice. First, we formulate the problem using a network utility maximization (NUM) framework, decompose the problem, and provide a distributed solution. Then, based on the structure of the NUM solution, we design a system called MicroCast, and we implement a prototype as an Android application. We provide both simulation results of the NUM solution and experimental evaluation. We demonstrate that the proposed approach brings significant performance benefits (namely, faster download on the order of the group size) without battery penalty.

Designing a communication protocol for sensor networks often involves obtaining the right trade-off between energy efficiency and end-to-end packet error rate. In this article, we show that network coding provides a means to elegantly balance these two goals. We present the design and implementation of SenseCode, a collection protocol for sensor networks—and, to the best of our knowledge, the first such implemented protocol to employ network coding. SenseCode provides a way to gracefully introduce a configurable amount of redundant information into the network, thereby decreasing end-to-end packet error rate in the face of packet loss. We compare SenseCode to the best (to our knowledge) existing alternative and show that it reduces end-to-end packet error rate in highly dynamic environments, while consuming a comparable amount of network resources. We have implemented SenseCode as a TinyOS module and evaluate it through extensive TOSSIM simulations.

Video streaming is one of the increasingly popular, as well as demanding, applications on smartphones today. In this paper, we consider a group of smartphone users, within proximity of each other, who are interested in watching the same video from the Internet at the same time. The common practice today is that each user downloads the video independently using her own cellular connection, which often leads to poor quality.

In this work, we are interested in a scenario where a group of smartphone users, within proximity of each other, are interested in watching the same video at the same time. The default operation today is that each user with a cellular connection downloads the video independently from the server. However, each phone's individual cellular connection may not be sufficient for providing high video quality.

This paper experimentally studies the reliability and delay of flooding based multicast protocols for a sniper detection application. In particular using an emulator it studies under which conditions protocols based on network coding deliver performance improvements compared to classic flooding. It then presents an implementation of such protocols on mobile phones.

This paper considers function computation in a network where intermediate nodes perform randomized network coding, through appropriate choice of the subspace codebooks at the source nodes. Unlike traditional network coding for computing functions, that requires intermediate nodes to be aware of the function to be computed, our designs are transparent to the intermediate node operations.

In a significant class of sensor-network applications, the identities of the reporting sensors constitute the bulk of the communicated data, whereas the message itself can be as small as a single bit - for instance, in many cases, sensors are used to detect whether and where a certain interesting condition occurred, or to track incremental environmental changes at fixed locations. In such scenarios, the traditional network-protocol paradigm of separately specifying the source identity and the message in distinct fields leads to inefficient communication. This work addresses the question of how communication should happen in such identity-aware sensor networks. We calculate theoretical performance bounds for this type of communication, where 'performance' refers to the number of transmitted bits. We propose a communication protocol, where the identity and message of each source are specified jointly using subspace coding. We show through analysis and simulation that our protocol's performance is close to optimal and compare it to the performance of a traditional protocol, where identity and message are specified separately.

In a significant class of sensor-network applications, the identities of the reporting sensors constitute the bulk of the communicated data, whereas the message itself can be as small as a single bit - for instance, in many cases, sensors are used to detect whether and where a certain interesting condition occured, or to track incremental environmental changes at fixed locations. In such scenarios, the traditional network-protocol paradigm of separately specifying the source identity and the message in distinct fields leads to inefficient communication. This work addresses the question of how should communication happen in such identity-aware sensor networks. We reexamine the traditional source-identity/message separation and propose a scheme for jointly encoding the two. We use this to develop a communication method for identity-aware sensor networks and show it to be energy efficient, simple to implement, and gracefully adaptable to scenarios frequently encountered in sensor networks - for instance, node failures, or large numbers of nodes where only few are active during each reporting round.

We consider the problem of minimizing delay when broadcasting over erasure channels with feedback. A sender wishes to communicate the same set of mu messages to several receivers over separate erasure channels. The sender can broadcast a single message or a combination (encoding) of messages at each timestep. Receivers provide feedback as to whether the transmission was received. If at some time step a receiver cannot identify a new message, delay is incurred. Our notion of delay is motivated by real-time applications that request progressively refined input, such as the successive refinement of an image encoded using multiple description coding. Our setup is novel because it combines coding techniques with feedback information to the end of minimizing delay. It allows Theta(mu) benefits as compared to previous approaches for offline algorithms, while feedback allows online algorithms to achieve smaller delay than online algorithms without feedback. Our main complexity results are that the offline minimization problem is NP-hard when the sender only schedules single messages and that the general problem remains NP-hard even when coding is allowed. However we show that coding does offer delay and complexity gains over scheduling. We also discuss online heuristics and evaluate their performance through simulations.

In networks that employ network coding, two main approaches have been proposed in the literature to allow the receivers to recover the source information: (i) use of coding vectors, that keep track of the linear combinations the received packets contain, and (ii) subspace coding, that dispenses of the need to know the linear combinations, since information is conveyed from the choice of subspaces alone. Both these approaches impose the strong requirement that all source packets get potentially combined. We here present a third approach that relaxes this assumption, and is thus not a special case from either of the previous two. This relaxation allows to employ compressed coding vectors to efficiently convey the coding coefficients, without altering the operation of intermediate network nodes. We develop optimal designs for such vectors.

We present ConfErr, a tool for testing and quantifying the resilience of software systems to human-induced configuration errors. ConfErr uses human error models rooted in psychology and linguistics to generate realistic configuration mistakes; it then injects these mistakes and measures their effects, producing a resilience profile of the system under test. The resilience profile, capturing succinctly how sensitive the target software is to different classes of configuration errors, can be used for improving the software or to compare systems to each other. ConfErr is highly portable, because all mutations are performed on abstract representations of the configuration files. Using ConfErr, we found several serious flaws in the MySQL and Postgres databases, Apache web server, and BIND and djbdns name servers; we were also able to directly compare the resilience of functionally-equivalent systems, such as MySQL and Postgres.