Abstract

Over the years, Earthquake has been one of the leading cause of life loss and destruction of properties. Man-made structures like buildings, roads, dams and bridges are severely affected due to earthquakes. Bhuj earthquake of 2001 killed nearly 15,000 people and damaged around 3.5 lakh buildings. The major contributors to the loss of so many buildings were the poor construction and maintenance practices. To address the deficiencies that lead to such catastrophic damage, major reforms have been introduced to the Indian Standard codes and numerous studies have been conducted to prescribe methods to curtail the damages further. These studies can broadly be classified into earthquake prediction studies and earthquake preparedness studies. Earthquake prediction studies can be used to minimize the loss to life. But over the years, the research on earthquake prediction hasn’t provided satisfactory results, and Earthquake prevention looks like a far-fetched dream. Hence, earthquake preparedness is of utmost importance to minimize earthquake damage. Earthquake preparedness involves the construction of new buildings in compliance with the latest building construction codes of the country as well as fixing the deficiencies of the existing buildings. Hence, identifying the vulnerabilities in a building is a very crucial step towards earthquake risk mitigation. When the vulnerable buildings of a region are identified, the seismic risk of the region can be determined, and actions to mitigate these risks can be taken. But the identification of vulnerable buildings in a region is a very resource and time-intensive process. A quick method to identify vulnerabilities in the buildings of a particular region is Rapid Visual Screening (RVS). This method can be used to find common vulnerabilities in the buildings, and the entire process of RVS of a building can be completed within 30 minutes by a single, trained person. But many buildings may require a detailed investigation in order to assess its safety. Detailed investigation requires the modelling and analysis of the building in finite element software. This requires the time and energy of people who are proficient in such software along with powerful machines to run such software. These problems make detailed investigation, time and resource-intensive process as well as very expensive. Hence, there is a need for an efficient and effective method to perform this task.

Abstract

The problem of multi-objective decision making has been an important one in the field of database management systems. Establishing dominance relationships among different data points has useful applications in industries such as tourism & sports. Given a dataset D with c points, a point pi dominates another point pj if it is better than or equal to pj in all the dimensions in space S and better than pj in at least one dimension. The set of such points which dominate all other data points in the dataset is referred to as the skyline set of points. The skyline operator suffers from the curse of dimensionality which means as the number of dimensions increase the number of skyline points become large and difficult to compute. With data growing at exponential rate in today's information age traditional algorithms designed for skyline query computation over limited resource standalone computers will not fare too well. Scaling up systems in terms of computing power & memory has traditionally proven to be expensive. Therefore, there is a need to leverage a cluster of cheap commodity hardware to efficiently compute skylines of large datasets. Thus the data needs to be divided into multiple smaller datasets which can be used to compute individual skylines across the smaller datasets and merged to form the overall skyline of the entire dataset by using a divide and conquer approach. The curse of dimensionality is even more amplified on large datasets hence making the problem of computing skyline even more complex, thus motivating us to use techniques which do not require us to eliminate points already in the computed skyline by points encountered afterwards. We achieve this by approaching points along the Z-order curve which provides us a monotonic access order of data points thus identifying skyline points earlier. The Z-address of a point in multi-dimensions is simply calculated by interleaving the binary representations of its coordinate values. In this thesis, we address the problem of efficient computation of skyline for large datasets using the Map-Reduce programming paradigm. Our approach is to map multi dimensional data to one dimension while preserving the locality of the data points using Z-order curve and exploiting the sort phase of a typical Map-Reduce job to receive these points in the Reduce job automatically sorted in the requisite order. Our results show a reduction of up to 31% in execution time for skyline computation over a Map-Reduce framework using our algorithm as compared to existing methods.

Abstract

Automated Question Answering (AQA) is an attractive variation in search engines where the QA system automatically returns a passage, which answers the user’s question instead of providing several links and letting the user explore for an answer. These AQA systems are an integral part of search engines and voice assistants such as Amazon Alexa, Apple Siri, Microsoft Cortona, Google Assistant, etc., Due to the recent growth in AQA systems, the research in this field has seen a tremendous improvement. Ranking the passages is an important step in an AQA system, where the candidate passages are identified for a question and scored as likely to contain an answer. To explore the various practical approaches for this problem, Microsoft has released a large scale dataset called as Microsoft MAchine Reading & COmprehension (MS MARCO) in 2016. Using the subset of the MS MARCO dataset, it has also organized a challenge called Microsoft AI Challenge India (MSAIC) to provide a platform for researchers and AI practitioners across the country to contribute to this area. In this thesis, we propose different approaches based on biLSTM that we used during our participation in Microsoft AI Challenge India. We show that our best model, based on the biLSTM encoder with coattention mechanism (which we call as coattention encoder), has significantly outperformed the shared baselines and benchmark models. In the second part of this thesis, we discuss and address the two limitations of the coattention encoder. We first extend the coattention encoder to n-grams to make it capture the local context. We further improve the coattention encoder by proposing a simple attention pooling using the query to refine the passage’s representation. We perform all our methods on the MS MARCO passage ranking task and demonstrate that these two methods yields a relative increase of 8.04% in Mean Reciprocal Rank (MRR@10) compared to the naive coattention encoder. It is the best non-transformer model and is competitive to the BERT base while only having fewer parameters. Additionally, when compared to BERT, our model has a very low ranking latencies.

Abstract

Blockchain technology has been one of the most disruptive innovations in the last decade. Blockchain is a decentralized, append-only trustworthy distributed database. It was invented to build Bitcoin – to democratize financial systems, in particular, currency. However, soon the researchers learned the power of blockchain to build smart contracts, programs that can execute and validate contracts, which otherwise were traditionally done through legal contracts. Since then, there are a plethora of applications has been envisioned using blockchain. At the backend, miners are the ones who maintain the blockchain in anticipation of fair rewards. Bitcoin offers two types of rewards for the miners: (i) block rewards that halve every four years and (ii) transaction fees. Halving block rewards is done by making currency inflation-free. Transaction fees are optional. This can lead to stagnation of transactions and unprofitability for miners. Another challenge with Bitcoin or many other blockchains, to build real-world applications, is performance. Many times, improving performance in blockchain leads to different security issues. In this thesis, we address both research challenges: (C1), how to configure the Bitcoin reward structure to maintain its ecosystem healthily as well as inflation-free. (C2), how to design a blockchain that is performant as well as secure. Towards C1, we propose BitcoinF, a simple modification to Bitcoin rewards that ensures fairness to the miners. We analyze it game theoretically. Towards C2, we propose QuickSync, a novel blockchain protocol that is faster than most of the secure blockchain protocols such as Bitcoin and Ourborous and retains all the security properties. We do security analysis for different attacks such as Sybil attacks and prove the robustness of QuickSync towards the attacks we studied

Abstract

Machine learning methods have found applications in various branches of healthcare including disease diagnosis, disease progression and outcome prediction, drug design, etc. The computational nature of such methods democratizes medical expertise by enabling access to diagnosis and outcome prediction in the absence of a medical expert. In this thesis we present two different applications of AI in health care. We present LPGNet an efficient deep learning based model for diagnosis of Parkinson’s disease from Gait. Next, We present the statistical analysis and models built for severity and mortality prediction in hospitalized Indian COVID-19 patients. In the first part of the thesis we present LPGNet an efficient method for diagnosis of Parkinson’s Disease (PD) from gait. PD is a chronic and progressive neurological disorder that results in rigidity, tremors and postural instability. There is no definite medical test to diagnose PD and diagnosis is mostly a clinical exercise. Although guidelines exist, about 10-30% of the patients are wrongly diagnosed with PD. Hence, there is a need for an accurate, unbiased and fast method for diagnosis. In this study, we propose LPGNet, a fast and accurate method to diagnose PD from gait. LPGNet uses Linear Prediction Residuals (LPR) to extract discriminating patterns from gait recordings and then uses a 1D convolution neural network with depth-wise separable convolutions to perform diagnosis. We also undertake an analysis of various cross-validation strategies used in literature in PD diagnosis from gait and find that most methods are affected by some form of data leakage between various folds which leads to unnecessarily large models and inflated performance due to overfitting. The analysis clears the path for future works in correctly evaluating their methods. LPGNet achieves an AUC of 0.91 with a 21 times speedup and about 99% lesser parameters in the model compared to the state of the art which would enable the method to be implemented in embedded systems to enable automated Parkinson’s diagnosis to be cheap and widely accessible. In the second part of the thesis we present the statistical analysis to understand and build models for severity and mortality prediction in hospitalized Indian COVID-19 patients. The clinical course of coronavirus disease 2019 (COVID-19) infection is highly variable with the vast majority recovering uneventfully but a small fraction progressing to severe disease and death. Appropriate and timely supportive care can reduce mortality and it is critical to evolve better patient risk stratification based on simple clinical data, so as to perform effective triage during strains on the healthcare infrastructure. We analyse a diverse set of 70 bio-markers including physiological, haematological and genomic parameters from 544 COVID-19 patients from New Delhi, India to identify significant markers and build risk stratification and mortality prediction models. An XGboost classifier yielded the best performance on risk stratification (AUC score of 0.83). A logistic regression model yielded the best performance on mortality prediction (AUC score of 0.92) when tested on a holdout test set of 115 patients. Examination of the data in comparison to a similar dataset with a Wuhan cohort of 375 patients was undertaken to understand the much lower mortality rates in India and the possible reasons thereof. The comparison indicated higher survival rate in the Delhi cohort even when patients had similar parameters as the Wuhan patients who died. Our analysis of medications administered points to usage of steroids early on which may have helped in the survival of severe patients. This study helps in identifying high-risk patients and provides insights into the low mortality rate in India

Abstract

Use of Immersive Virtual Reality(VR) stack is expanding, as the organizations explore alternative methods to reduce the cost of training. The effectiveness of VR simulators has demonstrated significant gains in the transfer of training for domains such as surgical training, flight training, military reconnaissance, search and rescue operations, etc. However, despite significant advancements, there exists a substantial gap to study human cognition and behavior and to also train them for demanding tasks like military combat and reconnaissance. Many of the military systems in use today are robotic with a limited degree of decision-making ability and interaction with the environment where making mistakes isn’t an option and therefore requires human intervention to operate and command remotely. Thus, teleoperators need to be instructed and trained prior to perform such operations efficiently and effectively. For a better understanding, and increasing VR’s usability, we aim to identify the relevant constituent components of immersive technologies, and differentiate their roles. In this thesis, we aim to determine the role of two determinants (goal-specific instructions and display fidelity) of spatial knowledge that one acquires during spatial navigation in VR. We conducted two experiments to disentangle the contribution of the two factors individually. Literature has demonstrated the relative importance of instructions on task performance and having goals can help in focusing attention on relevant information in the environment. However, it is also dependent on the nature of goal or activity that one is asked to accomplish. From a navigation perspective, the simplest of the goal is to return to the starting point to end the activity. In the first study, we included this goal in the instructions for one group while the other group were only instructed to explore the virtual environment and build understanding. Acquired topological spatial knowledge was assessed from the drawn sketch-maps. Our results showed that participants with goal-specific instructions found it difficult to recall one out of three spatial relationships (Object Positioning). These results indicate that the presence of additional goals in the instruction can impact the spatial memory of the operators and can be detrimental to task performance. HMD VR compared to DT, provides a 3-D experience during the virtual exploration, it offers a sense of complete immersion in the virtual world but literature has shown mixed findings depending on the applications. To study the role of the second factor, we conducted a pilot study followed by a full-scale study to examine how two common VR display modes, head-mounted display (HMD) and desktop (DT), would affect spatial learning when we restrict ambulatory locomotion in HMD. The findings and shortcomings of the pilot study helped in the refinement of experimental design, protocols, and methodology. The two display modes were compared on spatial learning, time taken to complete the virtual exploration, motion sickness, workload, sense of presence, and spatial orientation ability. Our results highlight the differential effect of DT vs. HMD VR depends on the nature of cognitive and functional behavior that the user employs. We explain these results in terms of deficient idiothetic information in non-ambulatory HMD and lesser sensory conflicts in desktop mode. This thesis sets up the foundation stones and provides future research directions that could be used in developing more robust VR experiments and applications. It is intended to be useful for VR experimenters as well as end-users. These results are relevant for developing customized and sustainable virtual reality-based human-computer interactions.

Abstract

Online learning problems are among the most abundant machine learning problems, such as advertisement selection, movie recommendation, and node or link prediction in evolving networks. In online learning, one must make online, real-time decisions and continuously improve performance with the sequential arrival of data. The multi-armed bandit (MAB) problem is one of the fundamental online learning problems that captures the classic exploration vs. exploitation dilemma. As per the applications’ requirements, new variants of MAB have been developed to tackle the problem. For example, strategic bidding in the online learning environment (MAB mechanism); availability of extra information to make the decision (Contextual MAB); selecting multiple arms instead of the single-arm (Combinatorial MAB); a varying set of available arms (Sleeping MAB), etc. We identify some of the problems and appropriately address them by proposing efficient algorithms. We provide theoretical guarantees for the performance of the proposed algorithms and showcase their efficacy through simulations. Following are the problems we address in this thesis: • Designing truthful contextual multi-armed bandits with the backdrop of sponsored search auction [3]. We tackle the drawbacks of the current state-of-the-art. • Propose an efficient algorithm for sleeping combinatorial multi-arm bandit problem when we consider general reward function (satisfying mild smoothness conditions) [2]. • We take a multi-arm bandit approach to subset selection under constraints in an unknown environment [46].

Abstract

This thesis adds to the knowledge of a chemical named entity detection and chemical role labeling. We present guidelines for chemical named entity detection in the Examples section of patents. In the process, we created the second-largest chemical NER dataset, the WEAVE corpus. We also developed methods for computational named entity recognition. We describe the models along with their features and gaps. We simultaneously extend the WEAVE corpus with chemical role labels in the Examples section of patents. Chemical role labels tend to change temporally as the reaction proceeds step by step. A chemical that has the product’s role label in the first step would have a reactant role label in the next step. This fluid nature of labeling the same chemical in a coherent reaction discourse multiple times is yet to be computationally modeled. The dataset developed in this thesis would be the first step in that direction. We evaluated the baseline performance of the WEAVE corpus using Deep Affix-base BiLSTM-CRF based neural architecture. The model achieves an F-score of 91.37%, near state-of-the-art result. We also compare the baseline performance with BioCreative-V CHEMDNER-patents corpus. It was found to be complementary in nature, hence could be used for transfer learning. In future work, we suggest how this corpus could build systems that could automatically select the chemical reaction from literature and infer the reaction rule for a noninteractive organic synthesis design system.

Abstract

Music is omnipresent and has existed since time immemorial. In general, individuals’ response to music, be it in terms of movements or listening strategies, tend to be dictated by features intrinsic to music such as rhythmic structure, genre, in addition to extrinsic factors such as mood states, socio-cultural norms, amongst others. However, at the same time, human differences in physical, psychological, and behavioural traits affect the decisions we make and the experiences we pursue in daily life. Individual differences can refer to gender, cognitive styles of thinking, personality, amongst others. In this thesis, we try to expand our understanding of how individual characteristics modulate our musical behaviour. The study and analyses presents various angles of approaching musical behaviour in an interdisciplinary manner, particularly in the areas of embodied music cognition and music consumption. We particularly investigate musical behaviour in two broader sub-domains: Active Music Engagement, which entails corporeal involvement and Passive Music Engagement, which involves listening to music. In the first part, we look at decoding individual differences via active music engagement. The paradoxical balance between universality and individuality in human motoric responsiveness to music makes it interesting to note that all the individuals’ traits are encoded in free dance movement. The study addresses this by identifying individual differences, specifically gender, cognitive styles, Big Five personality traits, and music preferences using free dance movements via our proposed Machine Learning model. We further demonstrate the robustness of the proposed model by testing its efficacy on another data set and further providing conclusive evidence about the learned models’ generalizability. The results of the study support theories of embodied music cognition and the role of bodily movement in musical experiences by demonstrating the influence of gender, personality, and music preferences on embodied responses to heard music. In the second study, which deals with passive engagement, we investigate out how individual differences, specifically personality, are associated with the kind of emotional experiences one seeks from lyrics. We chose to investigate lyrics specifically because they are under-explored in the field of music information retrieval despite them playing a crucial role in eliciting emotions. Firstly, we sought to identify emotional connotations of music based on lyrics following which we associate them with individual differences represented by personality traits. To this end, first we propose a novel deep learning architecture to identify emotional connotations of music based on lyrics, which we further compare with the existing deep learning and even traditional methods on relevant datasets and show state-of-the-art performance. Subsequently, we use this model to extract emotional preferences of individuals mined via online music streaming platforms and associate them with inherent personality traits. Our findings validate our theory that various types of emotions conveyed by songs have a unique relationship with individual traits. This study contributes to a better understanding of the relationship between broad personality dimensions and the emotional experiences one naturally seeks on online music streaming platforms. Both of our studies corroborate previous research in the field in addition to providing novel findings in both the aspect of musical behaviour being modulated by individual differences.

Abstract

Next generation wireless networks (5G and beyond) present the challenges of serving clients via channels which are not traditional point-to-point communication channels. The study of efficient high throughput communication techniques for broadcast channels, interference channels, multiple access channels, device-to-device (D2D) communication, all obtain relevance in the present wireless communication scenario. The technique of utilizing local storage (which has become quite affordable, thanks to advances in hardware design), either on the client’s device or in a nearby location, for aiding communication services on a broadcast channel was introduced formally in the landmark paper by Maddah-Ali and Niesen, under the title of Coded Caching. In this paper, it was shown that a combination of (a) carefully utilizing the local storage or cache available to individual clients, and (b) coded transmissions during the delivery phase, brings tremendous gains in the rate of delivery of information of a broadcast channel. To obtain this large gain in the rate, the coded caching schemes in the landmark paper and state of the art literature, require that the files at the server be divisible into a large number of parts (this number is called subpacketization). In fact, most schemes require the subpacketization to be growing asymptotically as exponential in some r th root of the number of users K, with r being some positive integer. On the other extreme, few schemes having subpacketization linear in K are known; however they require extremely large number of users to exist, or they offer only little gain in rate. Because of these reasons the state of the art coded caching schemes are impractical for applications, even though they are theoretically sound. The first objective of this thesis is to propose a new coded caching scheme with low subpacketization and moderate rate gains utilizing projective geometries over finite fields. The proposed scheme achieve the asymptotic subpacketization, which is subexponential in K or exponential in O((log K) 2 ) (thus improving on the root-of-K exponent) and has a small cache requirement. As a special case of our proposed scheme, we get a new linear subpacketization scheme, which has a more flexible range of parameters than existing linear subpacketization schemes. Extending our techniques, we also obtain low subpacketization schemes for other multi-receiver settings such as D2D communications and distributed computing. We show numerical comparisons of our scheme with existing popular and state-of-the-art schemes, and see that our schemes achieve lower and practically relevant levels of subpacketization in many situations at the cost of higher rate (or equivalently, a lesser caching gain). In another line of work, secretive coded caching was proposed in the literature, where the cache content and the transmissions are required to be such that each user can decode only the requested file by that user, and no information about other files. In non-secretive coded caching, it was shown in the literature that commonality among the user demands help in reducing the rate of communication. The direct adaptation of this result to secretive coded caching is not possible because of the presence of an unique key in each transmission. The second objective of this thesis is to propose an improved (reduced rate) perfect secretive coded caching scheme by exploiting commonality among the user demands. The issues addressed in this thesis are summarized as follows: • Based on the ideas from Projective geometries over finite fields, this thesis leverages the set containment and the subspace containment properties, to develop subexponential and linear subpacketization coded caching schemes for practical number of users and low cache sizes. • This thesis extends the low subpacketization coded caching scheme developed (in this thesis) for the broadcast channel to – Coded caching in D2D networks. – Distributed computing framework. vii • By studying the precise leakage properties of a modified scheme of a state of the art secretive coded caching scheme, this thesis develops an improved (reduced rate) secretive coded caching scheme by exploiting commonality of the user demands