Abstract

Stroke is one of the leading causes of death and disability in the world. Early detection of Stroke (both hemorrhagic and ischemic) is very important as it can ensure up to full recovery. Timely detection of stroke, especially ischemic stroke is difficult as the changes in abnormal tissue only become visible after the damage has already been done. The detection is even more difficult on CT scan compared to other imaging modalities but the dependence of a large fraction of population on CT, makes the need to find a solution to the problem even more imperative. Though the detection accuracy of radiologists for early stroke depends on various factors like experience, available technology, etc., earlier estimates put the accuracy around 10% [45]. Even with considerable advancement in CT technology the performance has still only increased to around 70% or thereabouts [21]. Any kind of assistance to radiologists which can improve their detection accuracy would therefore be much appreciated. This thesis presents a framework for automatic detection and classification of different types of stroke. We characterize stroke as a distortion in the otherwise contralaterally similar distribution of brain tissue. Classification depends on the severity of the distortion with hemorrhage and chronic infarcts exhibiting the maximum distortion and hyperacute stroke showing the minimum. The detection work on hemorrhagic stroke and early ischemic stroke has clinical value whereas the work on later stages of ischemic stroke has mainly academic use. The automatic detection approach was tested on a dataset containing 19 normal (291 slices) and 23 abnormal (181 slices) datasets. The algorithm gave a high recall rate for hemorrhage (80%), chronic (95%), acute (91.80%) and hyperacute (82.22%) stroke at slice level. The corresponding precision figures were 93.3%, 90.47%, 87.5% and 69.81% respectively. The performance of the system in a normal vs. stroke-affected scenario was 83.95% precision and 86.74% recall. The lower precision value in case of hyperacute scans is because of large number of normal slices with slight disturbances in contra-lateral symmetry being identified as stroke cases. We also present a novel approach for enhancement of early ischemic stroke regions using image-adaptive window parameters, to aid the radiologists in the manual detection of early ischemic stroke. The enhancement approach increased the average accuracy of radiologists in clinical conditions from around 71% to around 90% (p=0.02, two tailed student’s t test) with the inexperienced radiologists benefitting more from the enhancement. The average reviewing time of the scans was also reduced from about 9 to 6 seconds per slice. Out of the two approaches, automatic detection and enhancement, results show the enhancement process to be more promising.

Abstract

Parsing morphologically rich, free word order languages like Hindi, Czech, Turkish, etc. is a challenging task. Unlike English, most of the parsers for such languages have adopted the dependency grammatical framework, which is better suited for the task.(Shieber, 1985; Meluk, 1988; Bharati et. al, 1995). Due to the availability of manually annotated corpora in recent years, data-driven dependency parsing has achieved considerable success. In this work, we explore the performance of three data-driven dependency parsers, Malt, MST and Partial parser, for Hindi dependency parsing. Our system uses a hindi corpus that has manually annotated dependency relations and automatically extracted features like POS, chunk, morphological information, etc. and does complete word level parsing. We achieved an accuracy of 86.5% Unlabeled Attachment Score(UAS) and 77.9% Labeled Attachment Score(LAS), which is the state-of-the-art for automatic hindi dependency parsing. We tried to improve the parsing accuracy by using two bootstrapping techniques, self-training and co-training. Self-training is a simple process of incorporating unlabeled data into the training data of the parser. It uses the parser’s own output on the unlabeled data to improve its performance. Co-training is a technique that uses the output of two or more parsers of comparable accuracy on unlabeled data to provide additional training data for each other. Considering the automatic hindi dependency parser as our baseline, we applied self-training by using a large raw corpus. We also performed co-training between the state-of-the-art parser(Malt), another parser with comparable accuracy(MST) and a parser which can learn from partial structures(Partial parser). We experimented with various parameters to select top sentences for bootstrapping like the average classifier score of sentence, product of classifier score, etc. We also extracted highly confident sub-parses using different techniques which were used to improve the performance of the Partial parser. Both self-training and co-training were performed using two different types of raw corpora, one from the same domain as the training and test data and another from a different domain. We present a comparative analysis of the effect of domain of raw corpora on bootstrapping techniques. We also compare the effect of using automatically parsed data via bootstrapping and additional gold data on parser accuracy. By bootstrapping, we achieved the best accuracy of 87.1% Unlabeled Attachment Score(UAS) and 78.8% Labeled Attachment Score(LAS), which is a significant improvement over the state-of-the-art accuracy for automatic parsing. Using highly confident sub-parses, we were able to improve the performance of the Partial parser from 79.85% UAS to 84.5% UAS, which is comparable to the state-of-the-art accuracy for automatic parsing.

Abstract

Translation is the task of transforming text from one natural language to another while preserving the meaning and structure of the original text. Current phrase-based models for statistical machine translation are extremely restrictive and lack the flexibility to incorporate necessary contextual features in discriminating translations for source phrases. In this thesis, we propose a contextual translation model trained using discriminative learning techniques. The contextual model is integrated with a phrase based decoder to translate a given input sentence. We show that a discriminative translation model trained using lexical and shallow syntactic contextual features results in better translations both in terms of improved lexical selection and sentence construction. The work in this thesis also discusses methods to extract syntactic grammars from an annotated treebank and create a syntactic lexicon to be used in training a Supertagger. We extend the set of contextual features used by incorporating rich syntactic contextual features in the form of “supertags” on the source and target side to improve translation performance. Evaluation is done using both empirical methods and subjective evaluation. Given that the current syntax-based models for machine translation use a cubic-time decoding algorithm to construct translations, the proposed method of incorporating rich syntactic information into phrase-based models while retaining the linear-time complexity of the decoding algorithm is one of the prime contributions of my thesis. Improvements obtained using contextual models are limited when faced with the problem of limited training data- a scenario commonly found in the case of Indian languages. It is possible to allocate certain resources to create parallel corpora resources either by hiring translators to manually translate documents from a source language or through increasingly fashionable crowdsourcing. The created resources should allow one to develop real-time machine translation systems with a satisfactory level of performance. In the second half of my thesis, I present a method to enrich the existing SMT models. These models use information from other sources like a transfer grammar from a rule-based machine translation system to improve the reordering quality while translating from English to Indian languages, which can easily be applied to other low-resource languages. Finally, I propose a novel method to predict what the learning curve of an SMT system would look like, to estimate how much parallel corpus is required in order to achieve a satisfactory translation performance. Given the enormity of the task to create parallel corpora for different Indian languages, we believe the learning curves predicted using our method can be used to develop parallel corpora resources efficiently.

Abstract

The immense growth of social media and web technologies have enabled users to create, share and exchange information in virtual communities and networks. The ease of usability has attracted considerable amount of users who collaborate on such social platforms. Content generated from such interactions and conversations contains a lot of information that could be of very good commercial and educational value. This necessitates the need to extract information from such conversations. Therefore, the task of information extraction has received lot of attention from industry as well as research community. However, the proportion of content generated is so huge that it leads to the problem of information overload and redundancy. These problems can be dealt appropriately by presenting crux of the content rather than the whole user-generated text. Automatic extraction of important topics, followed by summarization of the conversation may help in reducing considerable human effort to understand the content. Automatic text summarization is a well-known solution to the problem of information overload [3, 5]. Summarization of news articles has been explored from theory to practical models, but summarization of user-generated content has not been paid much attention to. This can be attributed to the noisy and unstructured text from social media which makes traditional Natural Language Processing (NLP) techniques difficult to apply. For our purpose, we have chosen real-time conversations that are composed of emails and chats from technical domain. Such conversations suffer from a bigger problem of data sparseness as the relations between various entities or objects of discussion and their attributes is not explicitly present. In our research, we try to address these problems by designing two different approaches, a supervised machine learning approach and a topic-focused approach. The machine learning approach models the sentence as a set of features and builds a trainable summarizer using these feature vectors. Thereafter, in topic-focused approach we find hidden topics of the conversation using external resources and bind our summary to these topics using a semantic sentence scoring model. In this thesis, we have mainly concentrated on extractive methods of summarization with sentence as a basic unit of the summary. This transforms our problem of summarization to “the selection of most representative and informative sentences”. We design a supervised machine learning framework where each sentence is represented as a set of features. Along with the frequency and heuristics based methods, we devise new features in the form of Discourse Marker and Sentiment Score to capture the relevancy of a sentence more accurately. We have used Rhetorical Structure Theory (RST) to describe the organization of text. Our feature set aims to capture the statistical, linguistic and sentimental aspects along with the dialogue structure of the conversation. Conversations are modeled using a set of features which represent sentences. Then we use various machine learning algorithms to build a trainable summarizer on these feature vectors. We use two different datasets and perform five fold cross validation of our approach. We conduct various set of experiments to show the robustness of our system. The inclusion of discourse marker and sentiment score feature leads to the generation of improved summaries. We apply different machine learning algorithms and show that our feature set is generic enough to learn from any of the algorithms. Our approach significantly outperforms the baselines on ROUGE F-scores. We take a step forward towards our study of analyzing instantaneous conversations by following an unsupervised approach. In contrast to our earlier approach where we propose new features to capture different aspects of a conversation, here we develop a topic-focused approach based on the characteristics of a conversation. The social media conversations attract a substantial number of participants and a single conversation tends to span a wide range of topics interspersed with irrelevant segments. Motivated by this, we detect the primary topic of discussion in the conversation and build a semantic word space using the discovered topics. However, because of the sparseness in data identifying the connected components, primary topic detection becomes a difficult task. There is not enough evidence and explicit relationship present in the text to link different entities and objects with their corresponding attributes. To tackle this problem we develop a two phase approach. In the first phase, we leverage topic modeling using web documents to find the primary topic of discussion in the conversation. Then, in the summary generation phase, we use relevance-based language modeling approach to score sentences depending on their association with the primary topic. For topic modeling, we have used Latent Dirichlet Allocation

Abstract

Computational Geometry emerged as an independent field in computer science in the last three decades. It deals with the design and analysis of data structures and algorithms for problems involving geometric objects. It has a lot of applications in Geographic Information Systems (GIS), computer graphics, databases, design in VLSI etc. A class of problems called geometric range searching problems deal with intersection of different geometric objects. In a standard geometric range searching problem, a set S of n geometric objects is to be preprocessed such that, given a query region q, efficiently report the objects or count the number of objects in S q, where S q is the space that belongs to both S and q. The range searching problems is a fundamental topic of computational geometry. It has been extensively studied and there exist many efficient solutions. In many applications, it is not sufficient to just report or count all the objects in S q. There is need for aggregation on the objects in S q to present more useful information. One such type of aggregation is geometric aggregation, where the aggregate function is geometric in nature. In this work, we consider the maxima query problem in two-dimensional space. Let S be the static set of n points in R2. A point p = (xp; yp) is said to dominate a point q = (xq; yq) if xp  xq and yp  yq. A point is maximal if it is not dominated by any other point in S. Given an orthogonal range query q, we need to efficiently report or count the maximal points of S q. We propose a linear size data structure for reporting the set of k maximal points among the points that lie in a given 3-sided orthogonal range unbounded on its right in O(log n+k) worst case query time. Using linear space, the data structure can support counting the number of maximal points in a 3-sided query range in O(log n) worst case time. This is the first data structure for counting maxima in a given orthogonal query range in plane. We propose data structures to support 4-sided maxima reporting in O(log n+k) worst case time using O(n log n) space, where k is the size of the output. Counting maxima in a 4-sided orthogonal query rectangle can be supported in O(log n) worst case time, using O(n log n) space. The techniques can be directly applied to solve rectangular visibility problem. We use the pointer machine model of computation.

Abstract

The tremendous innovation in technology in all spheres of life has led to an explosion in the amount of data that has been generated. However labeling this huge amount of data for supervised learning tasks is often quite expensive, time-consuming and require significant manual work. Active learning has emerged as a rapidly growing field of machine learning, which aims at reducing the labelling effort of the oracle (human expert) in acquiring informative training samples in domains where the cost of labeling is high. It has been applied in areas such as document classification, speech recognition, information extraction, etc to train classifiers by labeling minimal amount of data. However, most of the existing active learning strategies are computationally expensive as they select only a single sample for labeling by the oracle in every iteration. In this thesis we propose a framework for efficient batch mode active learning that is based on Reverse Nearest Neighbor queries. We develop an efficient sampling technique for active learning using the notion of influence sets based on Reverse Nearest Neighbor. The proposed sampling technique uses RNN queries to efficiently determine the informativeness of the sample based on a combination of factors such as their utility for the classifier, representativeness and the unique information that they provide. We extend this RNN based sampling technique to develop a framework that can label a batch of samples for efficient active learning. The framework selects a batch of samples such that the overall information gain for the classification model is maximized. This framework is made generic so that the underlying classification model can be changed as per the domain requirements. We also propose an extension to this framework to handle the multi-class classification task effectively. Experiments conducted on the UCI and Reuters datasets show that our sampling framework and its variants are highly competitive in terms of accuracy in comparison with popular active learning techniques. We also develop an interactive application for bookmarks classification as a proof-of-concept of the utility of our batch mode active learning framework in interactive applications. This bookmarking system provides an opportunity for users to interact with the system by labeling the bookmarks to help the system learn about their preferences. The RNN based batch mode active learning framework is used in the backend to generate informative bookmarks for the users to label. Experiments show that labeling informative bookmarks can help the system to achieve good accuracy even with few labeled samples thereby providing a better user experience.

Abstract

Majority of the world’s electricity demand at present is met by thermal power generating stations which purely run on conventional fossil fuels. Utilities are mainly dependent on these sources spending huge revenue to cater for the ever increasing electricity demand. This motivates utilities to schedule their generation according to load demand in the most economical way. Every thermal generating unit is characterized by its distinct incremental heat rate curve which governs the production cost. An optimal allocation of generation among the different power generating units can thus save considerable fuel input and cost. However, a greater amount of fuel cost can potentially be saved by extending this optimization procedure to decide which of these units would participate in the optimal allocation. In other words, it is significant to determine whether the unit must be ON/OFF. Committing thermal generating units among the available ones in this fashion is popular as Unit Commitment (UC). It is a vital economic optimization problem in power systems with the main criteria of meeting the load demand while accounting for all the unit operational constraints at the same time. On the other hand, optimally distributing the generation among the set of units already scheduled/committed for operation is known as Economic Dispatch (ED). Present day load demand is increasing at a pace which expansion in power system has hardly been able to keep up with. The power systems have thus been left stressed. This makes the system insecure and vulnerable. Hence system security also plays a major role in generation planning. Therefore, the problem of UC is not merely to commit adequate generation but distribute it with utmost attention towards system security for achieving smart grid of future. In this thesis, the evolutionary technique of Binary Particle Swarm Optimization (BPSO) is used in solving the UC problem. As for determining the optimal generation dispatch for the minimal cost, a complete Optimal Power Flow (OPF) is run for each hour’s commitment over the UC time horizon. Conventional OPF solves for the optimal dispatch adjusting to all the constraints like fixed bus voltage limits, line power flows, transformer tap positions etc., resulting in a secure solution. Apart from these concerns, it would be very useful to compute the loadability and safety margin of a given system. This is important as it can potentially avoid any system collapse and keep a measure of system security in true sense. The work proposed in this thesis makes attempt to find optimal unit commitment keeping system operation secure by calculating what is known as the global L-index. It is a mathematical tool which computes how far the current operating condition is from the collapse point. Inclusion of such a constraint would naturally ensure that the optimal solution does not violate system voltage security. The major contribution of the proposed work lies in investigating more number of feasible solutions because the replacement of hard limits on the bus voltages by the L-index criterion effectively loosens the search space. Thus, experimentation is carried out by calculating the L-index while relaxing the fixed voltage limits on all or a few buses at the same time. A consistent raise in fuel costs and fast depletion of fossil fuels, growing awareness towards environmental protection has paved the path towards utilization of renewable energy sources for power production. Hence renewable energy sources are used and deployed in power systems with more eagerness today. However, they impose a considerable impact on the system owing to the intermittent nature of availability. The problem of UC becomes more complex with the integration of renewable sources given the dissimilarities in behavioural and functional constraints with respect to conventional thermal generation units which need to be addressed as renewable generation will be integral part of smart grid. This thesis aims at solving a voltage security constrained UC with the inclusion ofrenewable energy sources into the system. In the present work inclusion of one such source like Solar into the system has been considered. A Solar Thermal Power Plant (STPP) of suitable capacity is modelled with Concentrated Solar Power (CSP) technology of parabolic trough type. The standard dimensions of parabolic trough from a commercially installed power plant are considered for modelling. The optimal generation dispatch function (OPF) solution, which is an inherent exercise within the master problem of unit commitment, also accounts for the intermittent output from solar generator present in the same grid. The proposed method is validated by implementing it on the standard IEEE 14 and 30 bus test systems. One of the thermal generating units in standard test system is replaced by a solar thermal unit and the UC schedule is determined for the remaining units. The UC results, illustrated both in

Abstract

The main objective of this research is to contribute to the understanding of the seismic performance of superstructure considering the complex dynamic interaction between superstructure, the pile foundation and the soil. As the dynamic response of the structure and the pile to large extent is inelastic, the primary focus is on studying the behaviour of superstructure by modeling the nonlinearities of soil, modeling the interface between pile and soil. To address this problem, a Finite Element Method is used to model soil structure interaction analysis of pile supported framed structures by programming in MATLAB R2009a using Direct Method. A parametric study is conducted to understand the pile soil behaviour (Soil Foundation Interaction (SFI)) by changing various parameters, like pile and soil modulus, pile length, pile diameter and number of piles of the pile group. In each case the response is converted to frequency domain to understand the shift in frequency. Further, an attempt has been made to understand that complex behaviour of Soil Foundation Structure Interaction (SFSI). For that purpose, a 5 storey pile supported frame structure is modeled and its nonlinear behaviour under strong earthquake excitations is studied. A comparison of linear and nonlinear responses and the effect / significance of soil inelasticity on the structural response are commented. A SFSI system is modeled by considering the nonlinearity at the interface of the soil and pile. For that purpose, an interface element is used to model the interface between pile and soil. Parametric study has been carried out to know the response of pile with and without interface element and also to know the response of pile supported framed buildings with and without interface element. Besides this, the change in response of a high rise structure when a group of adjacent pile supported structures are present under seismic excitation is also studied (Structure Soil Structure Interaction (SSSI)). Different case studies are considered, namely 1. the group effect of structures supported on piles are considered (like group of two identical structures, group of three identical structures and group of three different structures), 2. the effect of variability in structure height is considered (like 5 storey structure, 10 storey structure and 15 storey structure) and 3. the effect of variability in structure shape is considered. For each case, the SSSI response is compared with the conventional fixed base response to understand the significance of SSSI. Few quantitative conclusions as mentioned below are made out of this study by commenting the significance of each behaviour (free field over SFI, SFSI over fixed base analysis, SSSI over fixed base analysis). The presence of soil and foundation in a SSI make a considerable change in response of the structure with a shift of natural period of the system. A peculiar behavior in the stress state of pile is observed for both elastic and inelastic soils, this behavior is because of Soil resistance acting downwards along the pile shaft because of an applied transient load. Repeated dynamic contacts of soil and pile is observed for both SFSI and SFI, this is because of the lateral compression of soil leading to formation of gap between pile and soil. The behavior on stress state of pile is very much different for the case of analysis with Interface elements. SSSI effects have been found to be important, when a group of identical structures with same dynamic characteristics are present,. The middle structures are attracting more displacements because of trapping of seismic waves. In case of group of structures with variable height, while considering SSI there is a decrease in response for 15 storey structure when compared to 10 storey structure which is not observed in fixed base system. In case of response of structures of variable shape the top floors will attract more displacement because of reduced stiffness on top floors but in conventional fixed base case opposite behavior is observed. So it has been recognized from this study that a reasonable seismic analysis for high rise buildings supported on pile foundations is needed to produce a safe and economic design.

Abstract

One of the accessory proteins of SARS coronavirus, 3a, is a 274 residue long protein with three transmembrane (TM) domains. This protein has been suggested to oligomerize to form tetrameric ion channels, which assist viral release from host cells. However, the atomistic details of the three-dimensional structure of the protein are still not available. In this study, molecular dynamics (MD) simulations have been used to understand the conformational properties of the three transmembrane domains (TM1, TM2 & TM3), and to model a structural model of the bundle formed from these domains. Extensive MD simulations have been performed in implicit bilayer environment with four different thicknesses, and in presence of explicit lipid bilayer environments (DLPE, DMPC, POPC, DPPC). The tilt angles to overcome the hydrophobic mismatch obtained at these eight different models were found to be slightly different from each other, but the inherent properties of the TM domains themselves were independent of the lipid bilayer model. TM1 domain was found to have a kink around Pro42 consistently in all the simulations. The conformational dynamics of each of the three TM domains were understood using tilt angle, kink angle, intra- and inter-molecular hydrogen bonding, and solvent accessible surface area. Using the stable conformations obtained using the simulations on the individual α-helices, a method by which potential three-dimensional structural models for the bundles is proposed. The surface areas of overlapping residues of possible bundles were calculated, and four different possible structures for the bundle are proposed. The best model structure comprising all three alpha helices was subjected to MD simulations to examine its quality. The TM bundle was found to form a compact and stable structure with significant intermolecular interactions. Further analysis indicates that residues from the TM2 and TM3 domains are likely to line the pore of the ion channel, which is in good agreement with a recent experimental study on the ion channel activity of SARS 3a protein.

Abstract

The earthquake resistant design of structures according to the existing design philosophy aims to ensure that during their lifetime, the structures resist the maximum possible earthquake without collapse. This philosophy ignores the fact that for the allowable damage at the end of the design life, it is necessary to account for the seismicity of the area in a comprehensive manner i.e., in terms of magnitudes with proper spatial distributions around the site. It is also necessary to estimate damage due to each of these events and also cumulative damage till date. Such a procedure is proposed in the report and it is applied to heritage structures which stand for longer duration compared to ordinary buildings. As a case study, the proposed methodology is used to evaluate damage to 3 Heritage structures,India gate and Qutb Minar located in Zone IV and Golconda fort located in Zone II according to IS 1893-2002 .Proposed approach is based on the estimation of expected number of earthquakes at each source with the help of Gutenberg-Richter relation and time-dependent hazard model and on the estimation of damage to SDOF structure during each of these events. The corresponding structural damage has been estimated by considering oscillator to be hysteretic in nature. The cumulative damage during the design life has been estimated by considering the most critical sequence of earthquake events and by considering event to event stiness and strength degradation for the oscillator.