Defense Notices

In real world environments, the speech signals received by our ears are usually a combination of different sounds that include not only the target speech, but also acoustic interference like music, background noise, and competing speakers. This interference has negative effect on speech perception and degrades the performance of speech processing applications such as automatic speech recognition (ASR), and hearing aid devices. One way to solve this problem is using source separation algorithms to separate the desired speech from the interfering sounds. Many source separation algorithms have been proposed to improve the performance of ASR systems and hearing aid devices, but it is still challenging for these systems to work efficiently in noisy and reverberant environments. On the other hand, humans have a remarkable ability to separate desired sounds and listen to a specific talker among noise and other talkers. Inspired by the capabilities of human auditory system, a popular method known as auditory scene analysis (ASA) was proposed to separate different sources in a two stage process of segmentation and grouping. The main goal of source separation in ASA is to estimate time frequency masks that optimally match and separate noise signals from a mixture of speech and noise. Three major aims are proposed to improve upon source separation in noisy and reverberant acoustic signals. First, a simple and novel algorithm is proposed to increase the discriminability between two sound sources by magnifying the head-related transfer function of the interfering source. Experimental results show a significant increase in the quality of the recovered target speech. Second, a time frequency masking-based source separation algorithm is proposed that can separate a male speaker from a female speaker in reverberant conditions by using the spatial cues of the sources. Furthermore, the proposed algorithm also has the ability to preserve the location of the sources after separation.

Finally, a supervised speech separation algorithm is proposed based on deep neural networks to estimate the time frequency masks. Initial experiments show promising results for separating sources in noisy and reverberant condition. Continued work is focused on identifying the best network training features and network structure that are robust to different types of noise, speakers, and reverberation. The main goal of the proposed algorithm is to increase the intelligibility and quality of the recovered speech from noisy environments, which has the potential to improve both speech processing applications and signal processing strategies for hearing aid technology.

Incomplete and numerical data are common for many application domains. There have been many approaches to handle missing data in statistical analysis and data mining. To deal with numerical data, discretization is crucial for many machine learning algorithms. However, few work has been done for discretization on incomplete data.

This research mainly focuses on the question whether conducting discretization as preprocessing gives better results than using a data mining method alone. Multiple Scanning is an entropy based discretization method. Previous research shown that it outperforms commonly used discretization methods: Equal Width or Equal Frequency discretization. In this work, Multiple Scanning is tested on C4.5 and MLEM2 on in- complete numerical data sets. Results show for some data sets, the setup utilizing Multiple Scanning as preprocessing performs better, for the other data sets, C4.5 or MLEM2 should be used by themselves. Our secondary objective is to test which of the three known interpretations of missing attribute value is better when using MLEM2. Results show that running MLEM2 on data sets with attribute-concept values performs worse than the other two types of missing values. Last, we compared error rate be- tween C4.5 combined with Multiple Scanning (MS-C4.5) and MLEM2 combined with Multiple Scanning (MS-MLEM2) on data sets with different percentage of missing at- tribute values. Possible rules induced by MS-MLEM2 give a better result on data sets with "do-not-care" conditions. MS-C4.5 is preferred on data sets with lost values and attribute-concept values.

Our conclusion is that there are no universal optimal solutions for all data sets. Setup should be custom-made based on the data sets.

Time and again, fiber optic medium has proved to be the best means for transporting global data traffic which is following an exponential growth trajectory. High bandwidth applications based on cloud, virtual reality and big data, necessitates maximum effective utilization of available fiber bandwidth. To this end, multicarrier superchannel transmission systems, aided by robust digital signal processing both at transmitter and receiver, have proved to enhance spectral efficiency and achieve multi tera-bit per second data rates.

With respect to transmission sources, laser technology too has made significant strides, especially in the domain of multiwavelength sources such as quantum dot passive mode-locked laser (QD-PMLL) based optical frequency combs. In the present research work, we characterize the phase dynamics of comb lines from a QD-PMLL based on a novel multiheterodyne coherent detection technique. The inherently broad linewidth of comb lines which is in the order of tens of MHz, make it difficult for conventional digital phase noise compensation algorithms to track the large phase noise especially for low baud rate subcarriers using higher cardinality modulation formats. In the context of multi-subcarrier Nyquist pulse shaped superchannel transmission system with coherent detection, we demonstrate through measurements, an efficient phase noise compensation technique called “Digital Mixing” which exploits the mutual phase coherence among the comb lines. For QPSK and 16 QAM modulation formats, digital mixing provided significant improvement in bit error rate (BER) performance.  For short reach data center and passive optical network-based applications, which adopt direct detection, a single optical amplifier is generally used meet the power budget requirements to achieve the desired BER.  Semiconductor Optical Amplifier (SOA) with its small form factor, is a low-cost power booster that can be designed to operate in any desired wavelength and most importantly can be integrated with the transmitter. However, saturated SOAs introduce nonlinear distortions on the amplified signal. Alongside SOA, the photodiode also introduces nonlinear mixing in the form of Signal-Signal Beat Interference (SSBI). In this research, we study the impact of SOA nonlinearity on the effectiveness of SSBI compensation in a direct detection OFDM based transmission system. We experimentally demonstrate a digital compensation technique to undo the SOA nonlinearity effect by digitally back-propagating the received signal through a virtual SOA, thereby effectively eliminating the SSBI. ​

The Chinese Postman Problem also known as Route Inspection Problem is a famous arc routing problem in Graph theory. In this problem, a postman has to deliver mail to the streets such that all the streets are visited at least once and return to his starting point. The problem is to find out a path called the optimal postman tour such that the distance travelled by the postman by following this path is always the minimum distance that has to be travelled to visit all the streets at least once. In graph theory, we represent the street system as a weighted graph whose edges represent the streets and the street intersections are represented by the vertices. A graph can be directed, undirected or a mixed graph. Directed and undirected edges represent the one way and the two way streets respectively. A mixed graph has both the directed and undirected edges.

The Chinese postman problem can be divided into several sub problems of which finding the minimum cost perfect matching is the critical part. For a directed graph, the minimum cost perfect matching of a bipartite graph has to be computed. For an undirected graph, the minimum cost perfect matching of a general graph has to be computed. There are different matching algorithms to compute the minimum cost perfect matching efficiently. In this project, I have understood and implemented four different matching algorithms used in computing an optimal postman tour, the Edmond’s Blossom Algorithm and a Branch and Bound Algorithm for the directed graph and the Hungarian Algorithm and a Branch and Bound Algorithm for the undirected graph. The objective of this project is to compare the performance of these matching algorithms on graphs of different sizes and densities."

Many online social networks are increasingly being used as information sharing platforms. With a massive increase in the number of users participating in information sharing, an enormous amount of information becomes available on such sites. It is vital to preserve user’s privacy, without preventing them from socialization. Unfortunately, many existing models overlooked a very important fact, that is, a user may want different information boundary preference for different information. To address this short coming, in this paper, I will introduce a ‘social circle’ model, which follows the concepts of ‘private information boundaries’ and ‘restricted access and limited control’. While facilitating socialization, the social circle model also provides some privacy protection capabilities. I then utilize this model to analyze the most popular social networks (such as Facebook, Google+, VKontakte, Flickr, and Instagram) and demonstrate the potential privacy vulnerabilities in some of these networking sites. Lastly, I discuss the implication of the analysis and possible future directions. 

Moral inclinations expressed in user-generated content such as online reviews or tweets can provide useful insights to understand users’ behavior and activities in social networks, for example, to predict users’ rating behavior, perform customer feedback mining, and study users' tendency to spread abusive content on these social platforms.  In this work, we want to answer two important research questions. First, if the moral attributes of social network data can provide additional useful information about users' behavior and how to utilize this information to enhance our understanding. To answer this question, we used the Moral FoundationsTheory and Doc2Vec, a Natural Language Processing technique, to compute the quantified moral loadings of user-generated textual contents in social networks. We used conditional relative frequency and the correlations between the moral foundations as two measures to study the moral break down of the social network data, utilizing a dataset of Yelp reviews and a dataset of tweets on abusive user-generated content. Our findings indicated that these moral features are tightly bound with users' behavior in social networks. The second question we want to answer is if we can use the quantified moral loadings as new boosting features to improve the differentiation, classification, and prediction of social network activities. To test our hypothesis, we adopted our new moral features in a multi-class classification approach to distinguish hateful and offensive tweets in a labeled dataset, and compared with the baseline approach that only uses conventional text mining features such as tf-idf features, Part of Speech (PoS) tags, etc. Our findings demonstrated that the moral features improved the performance of the baseline approach in terms of precision, recall, and F-measure.​

The non-ending growth of data traffic resulting from the continuing emergence of high-data-rate-demanding applications sets huge capacity requirements on optical interconnects and transport networks. This requires optical communication schemes in these networks to make the best possible use of the available optical spectrum per a single optical channel to enable transmission of multiple tens of tera-bits per second per a single fiber core in high capacity transport networks. Therefore, advanced modulation formats are required to be used in conjunction with energy-efficient and robust transceiver schemes. Important challenges facing these goals are the stringent requirements on the characteristics of optical components comprising these systems. Especially the laser sources. Laser phase noise is one of the most important performance-limiting factors in systems with high spectral efficiency. In this research work, we study the effects of different laser phase noise characteristics on the performance of different optical communication schemes. A novel, simple and accurate phase noise characterization technique is proposed. Experimental results show that the proposed technique is very accurate in estimating the performance of lasers in coherent systems employing digital phase recovery techniques. A novel multi-heterodyne scheme for characterizing the phase noise of laser frequency comb sources is also proposed and validated by experimental results. This proposed scheme is the first one of its type capable of measuring the differential phase noise between multiple spectral lines instantaneously by a single measurement. Moreover, extended relations between system performance and detailed characteristics of laser phase noise are also analyzed and modeled. The results of this study show that the commonly-used metric to estimate the performance of lasers with a specific phase recovery scheme, linewidth-symbol-period product, is not necessarily accurate for all types of lasers, and description of FM-noise power spectral profile is required for accurate performance estimation. We also propose an energy- and cost-efficient transmission scheme suitable for metro and long-reach data-center-interconnect links based on direct detection of field-modulated optical signals with advanced modulation formats, allowing for higher spectral efficiency. The proposed system combines the Kramers-Kronig coherent receiver technique, with the use of quantum-dot multi-mode laser sources, to generate and transmit multi-channel optical signals using a single diode laser source. Experimental results of the proposed system show that high modulation formats can be employed, with high robustness against laser phase noise and frequency drifting.

Resource limited embedded systems provide a great challenge to programming using functional languages.  Although these embedded systems cannot be programmed directly with Haskell, I show that an embedded domain specific language is able to be used to program them, and provides a user friendly environment for both prototyping and full development.  The Arduino line of microcontroller boards provide a versatile, low cost and popular platform for development of these resource limited systems, and I use these boards as the platform for my DSL research.

First, I provide a shallowly embedded domain specific language, and a firmware interpreter, allowing the user to program the Arduino while tethered to a host computer.  Shallow EDSLs allow a programmer to program using many of the features of a host language and its syntax, but sacrifice performance.  Next, I add a deeply embedded version, allowing the interpreter to run standalone from the host computer, as well as allowing the code to be compiled to C and then machine code for efficient operation.   Deep EDSLs provide better performance and flexibility, through the ability to manipulate the abstract syntax tree of the DSL program, but sacrifice syntactical similarity to the host language.   Using Haskino, my EDSL designed for Arduino microcontrollers, and a compiler plugin for the Haskell GHC compiler, I show a method for combining the best aspects of shallow and deep EDSLs. The programmer is able to write in the shallow EDSL, and have it automatically transformed into the deep EDSL.  This allows the EDSL user to benefit from powerful aspects of the host language, Haskell, while meeting the demanding resource constraints of the small embedded processing environment.

Energy harvesting from solar sources in an attempt to increase efficiency has sparked interest in many communities to develop more energy harvesting applications for renewable energy topics. Advanced technical methods are required to ensure the maximum available power is harnessed from the photovoltaic (PV) system. This dissertation proposed a new discrete-in-time extremum-seeking (ES) based technique for tracking the maximum power point of a photovoltaic array. The proposed method is a true maximum power point tracker that can be implemented with reasonable processing effort on an expensive digital controller. The dissertation presents a stability analysis of the proposed method to guarantee the convergence of the algorithm.

Two types of PV systems were designed and comprehensive frame work of control design was considered for a stand-alone and a three-phase grid connected system.

Grid-tied systems commonly have a two-stage power electronics interface which is necessitated due to the inherent limitation of the DC-AC (Inverter) power converging stage. However, a one stage converter topology, denoted as Quasi-Z-source inverter (q-ZSI) was selected that interface the PV panel which overcomes the inverter limitations to harvest the maximum available power.

A powerful control scheme called Model Predictive Control with Finite Set (MPC-FS) was designed to control the grid connected system. The predictive control was selected to achieve a robust controller with superior dynamic response in conjunction with the extremum-seeking algorithm to enhance the system behavior.

The proposed method exhibited better performance in comparison to conventional Maximum Power Point Tracking (MPPT) methods and require less computational effort than the complex mathematical methods.​

Remote Procedure Calls are an integral part of the internet of things and cloud computing. However, remote procedures, by their very nature, have an expensive overhead cost of a network round trip. There have been many optimizations to amortize the network overhead cost, including asynchronous remote calls and batching requests together.

In this dissertation, we present a principled way to batch procedure calls together, called the Remote Monad. The support for monadic structures in languages such as Haskell can be utilized to build a staging mechanism for chains of remote procedures. Our specific formulation of remote monads uses natural transformations to make modular and composable network stacks which can automatically bundle requests into packets by breaking up monadic actions into ideal packets. By observing the properties of these primitive operations, we can leverage a number of tactics to maximize the size of the packets.

We have created a framework which has been successfully used to implement the industry standard JSON-RPC protocol, a graphical browser-based library, an efficient byte string implementation, a library to communicate with an Arduino board and database queries all of which have automatic bundling enabled. We demonstrate that the result of this investigation is that the cost of implementing bundling for remote monads can be amortized almost for free, when given a user-supplied packet transportation mechanism.