Abstract

Various philosophers including Bertrand Russell and Ludwig Wittgenstein have shared the view that extra-linguistic reality and its nature are somehow linked with language. Given that language is a medium of capturing the objective truth around us, it is possible that the formal logic of language could also be inspired by how the world around us operates. Reality as we understand it, is a series of continuous interactions between objects. These interactions are loosely referred to us by as events. With constant rise in content generated every day, it has become arduous task to keep track of all the events happening in the real world. A major challenge information access tasks face is that they have to deal with colossal amounts of unstructured data, i.e., data whose format loosely implies its meaning described in a human understandable language. This makes understanding event mentions in text a key problem in Natural Language Processing. Event detection is an important Natural Language Processing and Information Retrieval problem, that has been extended to multiple domains and is now being tackled for languages other than English.In this thesis, we address the problem of event detection in Hindi. Hindi, like many other languages is an object-centric language that grounds a metaphysical system in objects, with sentence constructions mainly describing the properties (state) of a noun or the interaction between these nouns. If we were to use Hindi to define the nature of being, real world events would be formally captured as a series of interactions between nouns. We are further motivated by the fact that, in the past few years, the doctrine that actions are events has become an essential, and sometimes an unargued part of the received view in the philosophy of action, despite the efforts of a few philosophers to undermine the consensus. This task is an important one given the constant rise in volume of Hindi data being produced and consumed. A recent report shows that Hindi content consumption on the web is growing at a rate of 94% per year as compared to English’s 19%1 . Given a low resource language like Hindi, with abundance in unstructured text, we believe leveraging systems to automatically detect events in unstructured texts can enable significant advances in the performance of various information access tasks. The task of event detection in Hindi is relatively unexplored, partly because of syntactic features such as relative free word order and nature of verb adjuncts and partly due to the lack of structured data in Hindi. These properties of Hindi, increase the number of options available for expressing structured information, hence significantly complicating how chunks of text can be semantically interpreted.From the perspective of extracting latent information, defining events as semantico-syntactic objects allows for extracting relations between both events and entities, that are not explicitly mentioned. The extracted event information can then be readily consumed by downstream tasks such as Question Answering, Summarization and Knowledge Graph generation. The initial work presented in this thesis is directed towards tackling two major challenges. The primary being understanding how event mentions are captured in Hindi. This task, motivated by the interesting property of Hindi, where the syntax of the sentence is influenced by the event-semantics, becomes a compelling one. The observations of our studies are captured and formulated as two documents (annotation guidelines and annotation specifications) that allow accurate identification of event mentions in Hindi text. The second challenge we tackle is that of automating the task of event mention detection in Hindi texts. In order to kickstart exploration in the domain of machine learning approaches, we build and publish the largest event annotated dataset for Hindi. Our research sheds lights on the various possible methods to automate event detection in Hindi texts and compares the impact of techniques built on top of and without hand-crafted features. Machine learning techniques for automated event-detection have advanced fairly owing to the current rise in introduction and adoption of sophisticated neural architectures. However, these architectures are plagued by biases in opportunistically collected data sets for major languages. As a result of this, there is little to no progress in development of resources and solutions for low resource languages. Driven by this motivation, the final part of work attempts to tackle this by introducing an architecture, the MultiLingual Sequence Tagger (M-LiST) capable of training on a combination of four monolingual datasets and attaining state of the art performance for open domain event detection on the in three languages and current best performance for one of the languages.

Abstract

Stroke is the leading cause of severe chronic disability and the second cause of death worldwide with 15 million new cases and 50 million stroke survivors. The poststroke chronic disability may be ameliorated with early neuro rehabilitation where non-invasive brain stimulation (NIBS) techniques can be used as an adjuvant treatment to hasten the effects. However, the heterogeneity in the lesioned brain will require individualized NIBS intervention where innovative neuroimaging technologies of portable electroencephalography (EEGlabel) and functionalnear-infrared spectroscopy (fNIRS) can be leveraged for Brain State Dependent Electrotherapy (BSDE). In this thesis, we first present a theoretical model and a computational approach based on excitation–inhibition (E–I) balance hypothesis to objectively quantify the poststroke individual brain state using online fNIRS–EEG joint imaging. One of the key events that occurs following Stroke is the imbalance in local E–I (that is the ratio of Glutamate/GABA), which may be targeted with NIBS using a computational pipeline that includes individual “forward models” to predict current flow patterns through the lesioned brain or brain target region. Next, we give a proof of concept for our approach by using The Virtual Brain platform, which allows for individual, subject-specific models of the brain. We set up a pipeline that simulates tDCS on a realistic head model, and incorporated these electric field effects into the large scale brain network model to produce cortical activity. Using the pipeline set up, we investigate the effects of tDCS and E-I parameters on cortical activity, and validate it with experimental findings.

Abstract

The Standard Model of particle physics has been a successful theory in explaining the constituents of matter and the forces governing them (except gravity). However, some theoretical and experimental observations suggest that it’s not a complete theory of nature. Some measurements in B meson decays (such as the ratios R D , R D ∗ and R k ) and the value of muon anomalous magnetic moment, a μ deviate significantly from the Standard Model predictions. These anomalies can be explained through some Beyond Standard Model theories which predict the existence of a hypothesized particle called leptoquark. Leptoquarks can have O(TeV) masses and large couplings to third generation quarks. In this dissertation, we study single and pair-produced scalar leptoquarks, with left and right-handed couplings, that decay to top quarks and muons in 13 TeV proton-proton collisions at the LHC. We utilize HEPTopTagger to detect top quarks in the Monte Carlo event samples generated from Madgraph and passed on to Pythia for showering. Delphes is used to simulate detector effects. To discriminate our signal from the Standard Model background, we apply cuts on various physical quantities and calculate the 5σ discovery reach of the High Luminosity LHC (HL-LHC) at 3000 f b −1 . We find that the HL-LHC discovery reach at 5σ significance for the left-handed scenario is about 2 TeV. For the right-handed scenario, the discovery reach at 5σ is around 1.75 TeV. In the first chapter, we introduce the subject of high energy physics. In chapter two, we give a brief introduction to quantum field theory, the Standard Model of particle physics and some Beyond Standard Model (BSM) theories. Chapter three gives an overview of the experimental setup for our study. In chapter four, we introduce leptoquarks and describe some of its previous searches. Chapter five deals with the analysis of our study. In chapter six, we present the results and conclusions of the study

Abstract

In the recent times there has been a rapid growth in telecommunication industry. This leads to the demands of larger bandwidth and faster signal processing. However due to the limitations of electronic devices it is becoming increasingly difficult to cater to this ever increasing demand of bandwidth. It was later discovered that signal processing in optical devices is faster as compared to it’s electronic counterparts. This advantage of optical devices over electronic ones has attracted large number of researchers to study, understand and develop various all-optical devices. It is where the semiconductor ring laser provides a viable alternative as a key element to various faster optical devices. In this thesis we first introduce all-optical switching devices such as schmitt trigger and delta sigma modulator which are crucial part of signal processing. The effects of designer oriented parameters on various specifications like switching time, sampling speed and input amplitude are also studied and analyzed later on. The only parameters which can be readily varied externally are considered because those are the parameters a designer can tune. Such an analysis is essential for optimization of BDSM’S parameters and getting the desired sampling rate. In the last chapter an all-optical proteretic device is introduced which works in just opposite fash- ion to hysteretic devices and can enable us to sample signals at even higher speeds of modulation as compared to conventional scmhitt trigger. The reason for proteretic switch being faster is that it avoids delayed transition as found in conventional hysteretic switch. Lastly a novel method of modulation in the form of all-optical hysteretic proteretic switch is introduced. Rate equations of semiconductor ring laser is mathematically modeled in MATLAB.

Abstract

Web services are facing a constant threat from ever-increasing novel attacks. Attackers are continually coming up with new and unseen attacks whenever a vulnerability is found. Also, as the at- tackers have access to massive Botnets, these attacks are initiated immediately, before the interested party patches the vulnerability, to cause a denial of service on the targeted service. In this context, anomaly-based network intrusion detection techniques are valuable to protect the web services against these unknown attacks. Anomaly detection algorithms have gained significance to counter these unknown attacks, although they have high false alarms rates. As these algorithms can be effective on unknown attacks, web services can leverage them to defend against botnets. These algorithms need to be applied on the traffic which is handled by the scalable web services, in effect, there is a need for developing an intrusion detection system which is scalable with the web service it is protecting and also validate the alarms raised by these algorithms. Network-based anomaly detection is a well-mined area of research, with many projects that have produced algorithms to detect suspicious and anomalous activities. In this thesis, a bot detection architecture named BotD is proposed; this architecture provides a framework to plugin existing anomaly detection algorithms and the newer ones as they evolve. These installed algorithms analyze the incoming traffic and raise anomaly alarms when suspicious activity is detected. The architecture validates these alarms and takes appropriate actions. If the alarm validation turns out to be positive, the architecture marks the clients with such traffic patterns as bots and drops the traffic coming from them subsequently. In the case of false alarms, feedback is given to the algorithms to improve themselves. Moreover, BotD provides an interface for algorithm developers to monitor various metrics related to the algorithms plugged-in and takes care of scaling itself with the incoming traffic. Furthermore, in the thesis, we propose a network orchestration algorithm, used to transfer state across hosts using the SDN control layer. This proposed orchestration algorithm is used in this architecture to transfer plugged-in algorithms’ state efficiently across nodes while scaling. Finally, a simulation of the architecture is presented as a proof of concept for the proposed architecture.

Abstract

The construction of buildings is attained using force based design based on existing code provisions. Force based design has limitations as it only takes into account peak acceleration but not frequency content and duration of ground motion. A more reliable technique under consideration is energy based design which takes into account frequency content as well as duration of ground motion. Therefore it is essential to study the relation between different structure as well as earthquake characteristics with energy. In this study two types of structure, long period and short period, is considered because of the major difference in their response to earthquake. Therefore behavior of long and short time period structures is studied in terms of energy and damage induced when subjected to dynamic earthquake ground motion. Structures were modeled considering single degree of freedom following appropriate assumptions. A tool has been developed to study the behavior for long and short time period structures using MATLAB. Elastoplastic model was assumed for the nonlinear response. Response was calculated by solving dynamic equilibrium equations using Newmark’s step by step integration scheme. The tool developed is capable of performing time history analysis for getting response and this response was used to calculate imparted energy, hysteretic energy etc. over time duration of ground motion. Hysteretic energy is responsible for any damage induced in the structure because it accumulates only when the structure goes into non- linear region

Abstract

High power lasers are attractive sources for several applications. Because of good beam quality at all power level, distributed thermal loading, etc fiber lasers are gaining popularity in Directed Energy applications. There is a continued interest in scaling laser systems up to power levels suitable enough for DE systems. Beam combining systems for high average-power lasers began in the late 1960s and early 1970s. Early systems [1, 2] propagated the beams using heavy-water-cooled copper optics and open beam trains with commercial fans to provide fresh air. They have evolved in the intervening 40- plus years to include highly sophisticated gimbaled control systems with extremely high-reflectance uncooled optics and adaptive optics to compensate for less-than-ideal laser beams and for atmospheric distortions. Due to the limitations put by thermo optic effects gain saturation and scattering noise, laser beams cannot be amplified indefinitely. An alternative approach to building high power lasers is to use arrays of relatively lower power lasers. However, this approach requires that the beams from the array elements be combined to have the propagation characteristics of a single beam. Beam combining of laser arrays with high efficiency and good beam quality for power and radiance (brightness) scaling is a long-standing problem in laser technology. Semiconductor and fiber gain elements have attractive attributes for beam-combined sys- tems because of their ease in building in array formats, their high efficiency, and the ability to get near-diffraction-limited beams from individual elements. Beam brightness at the source is of limited importance when considering realistic DE propagation scenarios in turbulent atmospheres. Incoherent combination of the high-quality beams from fiber lasers can generate directional beams that are getting serious consideration for use as laser DE systems. For spatial beam combination where the laser beams propagate through the atmosphere, the atmo- spheric turbulence, pointing errors like platform jitter and boresight error play an important role in the performance of the combination technique. Different models of atmospheric turbulence have been explored. A functional implementation of the HAP model is shown in this thesis. Several architectures of the beam director for incoherent beam combination has been proposed. In these architectures, each of the beams are focused and targeted at a fixed point in space. Equations for inclined beam is derived using transformation rotation and the beam focusing equations have been modified to model real beams with non-ideal divergence. We have analyzed the criticality of the effect of atmospheric turbulence, beam wander, residual mechanical jitter and boresight error on the intensity and spot-size of incoherently combined laser beams.

Abstract

In the past decades, there have been multiple terrorist threats in the world and there is a lack of understanding of the effect of such dynamic loading on buildings. In order to understand the effect of blast loading, study of the dynamic behavior of structures is important. Current country codes do not provide proper guidelines to build a structure to mitigate the effects of such terrorist threats. Therefore, making it essential to study the behavior of structures subjected to impulsive loading. Reinforced walls are one of the first structural elements which are exposed to the loading in case of such events. Thus they become crucial to mitigate the effects of explosions. As part of this research, out of plane behavior of a vertical reinforced wall was studied when subjected to dynamic blast loading. Single degree of freedom system and finite element method were used to carry out numerical modeling of the reinforced wall with appropriate loading and structural approximations. As a part of this study, a tool was developed to implement the finite element method for a reinforced wall in MATLAB. 3D modeling of reinforced wall is done using 20-noded or 8-noded hexahedral elements. Rayleigh damping is used to calculate Damping matrix. The dynamic equilibrium equations are solved using Gaussian quadrature and Newmark’s direct integration scheme. The tool was used to analyze the behavior of reinforced walls under varying percentage steel, damping ratio, standoff distance, reinforcement diameter, charge weight, etc.

Abstract

Glaucoma is a major eye disease which when untreated, can gradually lead to irreversible loss in vision. The underlying causes are a loss of retinal nerve fibres (resulting in a thinning of the layer and enlargement of the optic cup) and peripapillary atrophy. Since these occur without any sign of symptoms in the initial stages, it is difficult to diagnose the disease in the early stages. Hence, the development of Computer-aided diagnostics (CAD) systems for early detection and treatment of the disease has attracted the attention of many medical experts and researchers alike. Optical Coherence Tomography (OCT) and Fundus photography are two widely used retinal imaging techniques for obtaining the structural information of the eye which helps to analyze and detect the diseases. Existing automated systems rely largely on fundus images for assessment of glaucoma due to their fast acquisition and cost. OCT images provide vital and unambiguous information for understanding the changes occurring in the retina, specifically related to the retinal nerve fiber layer and the optic nerve head which are essential for disease assessment. However, the high cost of OCT is a deterrent for deployment in screening at large scale. Hence, the focus of this thesis is to investigate the potential of integrating the two retinal imaging techniques which provide complementary information of the eye for developing automatic glaucoma screening system. Firstly, we propose a deep learning approach directly operating on 3D OCT volumes for glaucoma assessment which showed promising results, thus demonstrating the prominence of the highly discriminative features learnt from OCT for automated glaucoma detection. Next, we present a novel CAD solution wherein both OCT and fundus modality images are leveraged to learn a model that can perform a mapping of fundus to the OCT feature space. We show how this model can be subsequently used to detect glaucoma given an image from only one modality (fundus), thus enabling the automated screening operation to be executed on a large scale. The results show that fundus to OCT feature space mapping is an attractive option for glaucoma detection

Abstract

Language identification (LID) refers to the task of identifying the language being spoken by a speaker in a given speech signal. LID system acts as a front-end module for different applications such as multilingual automatic speech recognizer, multilingual dialogue system, and voice service systems. Notably, in an Indian scenario, where almost every state has a language of its own and every language having various dialects, development of a LID system becomes crucial. In literature, the standard acoustic features such as Mel-frequency cepstral coefficients, shifted delta cepstral coefficients and prosody (pitch, duration, and intonation) have been explored in the Indian context. These features cannot adequately discriminate the languages since most of the Indian languages have an overlapped set of phonemes. Even though Indian languages have an overlapped set of phonemes, due to phonotactic constraints, the characteristics of a particular sound unit differ in different languages which motivates us to explore phonetic features. In the literature, most of the works used senone based deep neural networks to extract phonetic features. We have used a joint acoustic model (JAM) which is trained using long short term memoryconnectionist temporal classification (LSTM-CTC) network to extract phonetic features. The JAM implicitly learns the language-specific information without any prior knowledge of the language. Hence, it is hypothesized that features extracted from JAM effectively discriminate the languages. These features are referred to as CTC features. The temporal variations in CTC features are modeled using the attention-based residual time-delay neural network to predict language ID. Attention mechanism aggregates frame-level features by selecting prominent frames through a parametrized attention layer. The performance of the proposed LID framework (CTC with attention-based RES-TDNN) has been evaluated on IIITH-ILSC database, which consists of 22 official Indian languages and Indian English. The proposed network outperformed other state-of-the-art methods with an equal error rate of 8.82%.