Ultra Low-Power Frequency Synthesizers for Duty Cycled IoT radios

Internet of Things (IoT), which is one of the main talking points in the electronics industry today, consists of a number of highly miniaturized sensors and actuators which sense the physical environment around us and communicate that information to a central information hub for further processing. This agglomeration of miniaturized sensors helps the system to be deployed in previously impossible arenas such as healthcare (Body Area Networks - BAN), industrial automation, real-time monitoring environmental parameters and so on; thereby greatly improving the quality of life. Since the IoT devices are usually untethered, their energy sources are limited (typically battery powered or energy scavenging) and hence have to consume very low power. Today's IoT systems employ radios that use communication protocols like Bluetooth Smart; which means that they communicate at data rates of a few hundred kb/s to a few Mb/s while consuming around a few mW of power. Even though the power dissipation of these radios have been decreasing steadily over the years, they seem to have reached a lower limit in the recent times. Hence, there is a need to explore other avenues to further reduce this dissipation so as to further improve the energy autonomy of the IoT node. Duty cycling has emerged as a promising alternative in this sense since it involves radios transmitting very short bursts of data at high rates and being asleep the rest of the time. In addition, high data rates proffer the added advantage of reducing network congestion which has become a major problem in IoT owing to the increase in the number of sensor nodes as well as the volume of data they send. But, as the average power (energy) dissipated decreases due to duty cycling, the energy overhead associated with the start-up phase of the radio becomes comparable with the former. Therefore, in order to take full advantage of duty cycling, the radio should be capable of being turned ON/OFF almost instantaneously. Furthermore, the radio of the future should also be able to support easy frequency hopping to improve the system efficiency from an interference point of view. In other words, in addition to high data rate capability, the next generation radios must also be highly agile and have a low energy overhead. All these factors viz. data rate, agility and overhead are mainly dependent on the radio's frequency synthesizer and therefore emphasis needs to be laid on developing new synthesizer architectures which are also amenable to technology scaling. This thesis deals with the evolution of one such all-digital frequency synthesizer; with each step dealing with one of the aforementioned issues. In order to reduce the energy overhead of the synthesizer, FBAR resonators (which are a class of MEMS resonators) are used as the frequency reference instead of a traditional quartz crystal. The FBAR resonators aid the design of fast-startup oscillators as opposed to the long latency associated with the start-up of the crystal oscillator. In addition, the frequency stability of the FBAR lends itself to open-loop architecture which can support very high data rates. Another advantage of the open-loop architecture is the frequency agility which aids easy channel switching for multi-hop architectures, as demonstrated in this thesis.

Geometry-aware analysis of high-dimensional visual information sets

Effrosyni Kokiopoulou

Over the past few decades we have been experiencing a data explosion; massive amounts of data are increasingly collected and multimedia databases, such as YouTube and Flickr, are rapidly expanding. At the same time rapid technological advancements in mobile devices and vision sensors have led to the emergence of novel multimedia mining architectures. These produce even more multimedia data, which are possibly captured under geometric transformations and need to be efficiently stored and analyzed. It is also common in such systems that data are collected distributively. This very fact poses great challenges in the design of effective methods for analysis and knowledge discovery from multimedia data. In this thesis, we study various instances of the problem of classification of visual data under the viewpoint of modern challenges. Roughly speaking, classification corresponds to the problem of categorizing an observed object to a particular class (or category), based on previously seen examples. We address important issues related to classification, namely flexible data representation for joint coding and classification, robust classification in the case of large geometric transformations and classification with multiple object observations in both centralized and distributed settings. We start by identifying the need for flexible data representation methods that are efficient in both storage and classification of multimedia data. Such flexible schemes offer the potential to significantly impact the efficiency of current multimedia mining systems, as they permit the classification of multimedia patterns directly in their compressed form, without the need for decompression. We propose a framework, called semantic approximation, which is based on sparse data representations. It formulates dimensionality reduction as a matrix factorization problem, under hybrid criteria that are posed as a trade-off between approximation for efficient compression and discrimination for effective classification. We demonstrate that the proposed methodology competes with state-of-the-art solutions in image classification and face recognition, implying that compression and classification can be performed jointly without performance penalties with respect to expensive disjoint solutions. Next, we allow the multimedia patterns to be geometrically transformed and we focus on transformation invariance issues in pattern classification. When a pattern is transformed, it spans a manifold in a high dimensional space. We focus on the problem of computing the distance of a certain test pattern from the manifold, which is also closely related to the image alignment problem. This is a hard non-convex problem that has only been sub-optimally addressed before. We represent transformation manifolds based on sparse geometric expansions, which results in a closed-form representation of the manifold equation with respect to the transformation parameters. When the transformation consists of a synthesis of translations, rotations and scalings, we prove that the objective function of this problem can be decomposed as a difference of convex functions (DC). This very property allows us to solve optimally our optimization problem with a cutting plane algorithm, which is well known to successfully find the global minimizer in practice. We showcase applications in robust face recognition and image alignment. The classification problem with multiple observations is addressed next. Multiple observations are typically produced in practice when an object is observed over successive time instants or under different viewpoints. In particular, we focus on the problem of classifying an object when multiple geometrically transformed observations of it are available. These multiple observations typically belong to a manifold and the classification problem resides in determining which manifold the observations belong to. We show that this problem can be viewed as a special case of semi-supervised learning, where all unlabelled examples belong to the same class. We design a graph-based algorithm, which exploits the structure of the manifold. Estimating the unknown object class then results into a discrete optimization problem that can be solved efficiently. We show the performance of our algorithm in classification of multiple handwritten digit images and in video-based face recognition. Next, we study the problem of classification of multiple observations in distributed scenarios, such as camera networks. In this case the classification is performed iteratively and distributively, without the presence of a central coordinator node. The main goal is to reach a global classification decision using only local computation and communication, white ensuring robustness to changes in network topology. We propose to use consensus algorithms in order to design a distributed version of the aforementioned graph-based algorithm. We show that the distributed classification algorithm has similar performance as its centralized counterpart, provided that the training set is sufficiently large. Finally, we delve further into the convergence properties of consensus-based distributed algorithms and we propose an acceleration methodology for fast convergence that uses the memory of the sensors. Our simulations show that the convergence is indeed accelerated in both static and dynamic networks, and that distributed classification in sensor networks can significantly benefit from them. Overall, the present thesis addresses a few important issues related to pattern analysis and classification in modern multimedia systems. Our solutions for semantic approximation and transformation invariance can impact the efficiency and robustness of classification in multimedia systems. Furthermore, our graph-based framework for multiple observations is able to perform effective classification in both centralized and distributed environments. Finally, our fast consensus algorithms can significantly contribute to the accelerated convergence of distributed classification algorithms in sensor networks.

EPFL2009

Adaptive Selection Problems in Networked Systems

Runwei Zhang

Networked systems are composed of interconnected nodes that work collaboratively to maximize a given overall utility function. Typical examples of such systems are wireless sensor networks (WSNs) and participatory sensing systems: sensor nodes, either static or mobile, are deployed for monitoring a certain physical field. In these systems, there are a set of problems where we need to adaptively select a strategy to run the system, in order to enhance the efficiency of utilizing the resources available to the system. In particular, we study four adaptive selection problems as follows. We start by studying the problem of base-station (BS) selection in WSNs. Base stations are critical sensor nodes whose failures cause severe data losses. Deploying multiple fixed BSs improves the robustness, yet this scheme is not energy efficient because BSs have high energy consumptions. We propose a scheme that selects only one BS to be active at a time; other BSs are kept passive and act as regular sensor nodes. This scheme substantially reduces the energy supplies required by individual BSs. Then, we propose an algorithm for adaptively selecting the active BS so that the spatially and temporally varying energy resources are efficiently utilized. We also address implementation issues and apply the proposed algorithm on a real WSN. Field experiments have shown the effectiveness of the proposed algorithm. We generalize the BS selection problem by considering both the energy efficiency of regular sensor nodes and that of BSs. In this scheme, a subset of active BSs (instead of only one) is adaptively selected and the routing of regular sensor nodes is adjusted accordingly. Because BSs have high fixed-energy consumptions and because the number of candidate subsets of active BSs is exponential with the number of BSs, this general BS selection problem is NP-hard. We propose a polynomial-time algorithm that is guaranteed, under mild conditions, to achieve a network lifetime at least 62% of the optimal one. Through extensive numerical simulations, we verify that the lifetime achieved by the proposed algorithm is always very close to the optimum. We then study the problem of scheduling the sparse-sensing patterns in WSNs. We observe that the traditional scheme of periodically taking sensing samples is not energy efficient. Instead, we propose to adaptively schedule when and where to activate sensors for sampling a physical field, such that the energy efficiency is enhanced and the sensing precision is maintained. The schedules are learnt from the temporal signal models derived from the collected measurements. Then, using the obtained signal models and the sparse sensing-measurements, the original signal can be effectively recovered. This proposed method requires minimal on-board computation, no inter-node communications and achieves an appealing reconstruction performance. With experiments on real-world datasets, we demonstrate significant improvements over both traditional sensing schemes and the state-of-the-art sparse-sensing schemes, particularly when the measured data is characterized by a strong temporal correlation. In the last part of the thesis, we discuss the sparse-sensing framework by exploiting the spatial correlations rather than the temporal correlations among the captured measurements. In this framework, application-specific utility functions can be employed. By adaptively selecting a small subset of active sensors for sensing, a certain utility is guaranteed and the efficiency of the sensing system is enhanced. We apply this framework both in static WSNs and participatory sensing systems where sensors move in an uncoordinated manner. Through extensive simulations, we show that our proposed algorithm enhances the resource efficiency.

EPFL2015

Localizing the Source of an Epidemic Using Few Observations

Brunella Marta Spinelli

Localizing the source of an epidemic is a crucial task in many contexts, including the detection of malicious users in social networks and the identification of patient zeros of disease outbreaks. The difficulty of this task lies in the strict limitations on the data available: In most cases, when an epidemic spreads, only few individuals, who we will call sensors, provide information about their state. Furthermore, as the spread of an epidemic usually depends on a large number of variables, accounting for all the possible spreading patterns that could explain the available data can easily result in prohibitive computational costs. Therefore, in the field of source localization, there are two central research directions: The design of practical and reliable algorithms for localizing the source despite the limited data, and the optimization of data collection, i.e., the identification of the most informative sensors. In this dissertation we contribute to both these directions. We consider network epidemics starting from an unknown source. The only information available is provided by a set of sensor nodes that reveal if and when they become infected. We study how many sensors are needed to guarantee the identification of the source. A set of sensors that guarantees the identification of the source is called a double resolving set (DRS); the minimum size of a DRS is called the double metric dimension (DMD). Computing the DMD is, in general, hard, hence estimating it with bounds is desirable. We focus on G(N,p) random networks for which we derive tight bounds for the DMD. We show that the DMD is a non-monotonic function of the parameter p, hence there are critical parameter ranges in which source localization is particularly difficult. Again building on the relationship between source localization and DRSs, we move to optimizing the choice of a fixed number K of sensors. First, we look at the case of trees where the uniqueness of paths makes the problem simpler. For this case, we design polynomial time algorithms for selecting K sensors that optimize certain metrics of interest. Next, turning to general networks, we show that the optimal sensor set depends on the distribution of the time it takes for an infected node u to infect a non-infected neighbor v, which we call the transmission delay from u to v. We consider both a low- and a high-variance regime for the transmission delays. We design algorithms for sensor placement in both cases, and we show that they yield an improvement of up to 50% over state-of-the-art methods. Finally, we propose a framework for source localization where some sensors (called dynamic sensors) can be added while the epidemic spreads and the localization progresses. We design an algorithm for joint source localization and dynamic sensor placement; This algorithm can handle two regimes: offline localization, where we localize the source after the epidemic spread, and online localization, where we localize the source while the epidemic is ongoing. We conduct an empirical study of offline and online localization and show that, by using dynamic sensors, the number of sensors we need to localize the source is up to 10 times less with respect to a strategy where all sensors are deployed a priori. We also study the resistance of our methods to high-variance transmission delays and show that, even in this setting, using dynamic sensors, the source can be localized with less than 5% of the nodes being sensors.

EPFL2018