Abstract

Landslides are the most common and devastating natural disaster in the Indian Himalayan region other than earthquakes. Landslides affect 15% of Indias land area which is around 0.49 million km2. It is estimated that 42% of all the landslide prone areas in the country fall in North East Himalayas, especially Darjeeling and Sikkim part of the Himalayas. The damage caused due to landslides is massive leading to loss of life, property including agricultural land. The studies on landslides have drawn worldwide attention due to the rapid increase in urbanization in many of these hilly regions and thus its increasing impact on socio-economic aspects. Thus there is a dire need for understanding landslides, estimating its occurrence potential (or modeling), and formulating strategies to minimize its impact. Some of the recent studies on Darjeeling-Sikkim Himalayas have largely looked at severity zonation, identifying rainfall thresholds and other related aspects. However, a comprehensive study of the entire region of Kalimpong has been largely left unattended. While, landslides can be considered to be either shallow or deep-seated failure of the soil mass, majority of the landslides in Kalimpong are categorized as shallow landslides. Such type of landslides are caused or reactivated or induced primarily by rainfall. In a recent report by Geological Survey of India (GSI, 2016) it identified that 75% of the landslide occurrences in this region during 2006-2013 was triggered by rainfall. Therefore, it is imperative to understand the relationship between landslide incidences and rainfall conditions during and before, primarily in the context of the Himalayan region in Kalimpong. There are mainly two types of methods to understand this relationship: physical and empirical. Phys-ical process models are based on numerical models which study the relationship between rainfall, pore water pressure, soil type, and volumetric water content that can lead to slope instability. Such study is usually site specific due to variation in soil properties. It is a challenge to extend this approach to large areas, as the extensive data that is required are usually not available. On the other hand, Empirical methods study the landslides that are caused by rainfall events both the heavy downpour that triggers instantaneous landslides and the low but continuous antecedent rain that destabilizes the slope and triggers landslide. In this present work, an empirical approach is taken for assessing the landslides in the Kalimpong region by considering the daily rainfall intensities. As such methods produce only binary results i.e, either landslides occur or do not occur, we have also adopted and evaluated probabilistic methods for this region. In addition, the study further explores the relationship between rainfall and landslide occurrences, by using a mathematical model to simulate the potential triggering conditions in Chibo, one of the most active landslide regions in Kalimpong. To validate and assess the empirical model, an IoT sensor based field observations were carried out. The installed sensors are Microelectromechanical Systems (MEMS) tilt sensor and volumetric water content sensors in the Chibo area. While the former measures the tilting angle of the instrument at shallow depths and hence the lateral displacement at the slope surface; the later measures the soil moisture levels. These were used to assess the model performance. Also, it was found that antecedent rainfall of 20 days or more is one of the major causes for rainfall induced landslides in this region. The results also signify that a rainfall intensity of 60-70 mm/day has the highest probability of landslide occurrence for the Kalimpong region. The results signify that to develop an operational early warning system without the need of a monitoring system threshold values would be equal to higher than empirical derived thresholds but not higher than hydrological thresholds. The work shows that Early warning systems can hence be designed based on these rainfall thresholds as the first line of action.

Abstract

An academic citation network can be viewed as an information graph where individual nodes contain rich textual information. With the current trend of open-access to most scientific literature, we presume that the full text of a scientific article (or paper) is a vital source of information which aids in various recommendation and prediction tasks concerning this domain. The textual content of a paper serves as a rich source of information about a node in the citation network graph. The first task we focus on is competing algorithm mining from research papers. As an example, consider the association rule mining problem in data mining or the object recognition problem in computer vision. There has been a lot of research on these particular ideas and several authors and research groups have come up with their novel solutions to tackle these problems. The authors would measure their proposed algorithms against current state-of-the-art. Given the textual content of a paper we propose to mine all the algorithms which are being compared against inside the paper of interest. As our first contribution, we develop an unsupervised mechanism leveraging natural language processing techniques to solve this problem. Recently representation learning using neural networks has been receiving much attention from both industrial and academic communities. Thanks to the advent of powerful GPUs and advanced techniques like dropout and rectified linear units, the interest in neural networks has been revived. The current state-of-the-art solution for the applications in the field of Natural Language Processing (NLP) and Information Retrieval (IR) is powered by neural network based representation learning models. We explore the similar theme of using text content from research papers and propose ways of improving search, retrieval and recommendation in academic citation networks. For example, two papers published between a short timespan would not neccessarily cite each other even if they cater to very similar ideas. As our second contribution we introduce Paper2vec: a novel two step neural network based approach to obtain vector representaions of scientific papers from an academic citation network by exploiting both its textual content and network (graph) properties. Our third contribution is on similar lines, where instead of papers we focus on authors inside the whole network. We specifically explore how we can exploit the content of papers written by an author to aid in possible collaboration prediction and propose Author2vec: a novel joint training approach using neural networks for the same. In the growing body of open access to research papers and journals, the set of algorithms and frame- works proposed in this thesis could act as a first step towards the use of full textual content for the aid of various prediction and recommendation tasks in academic citation networks.

Abstract

Apple Inc. unveiled the first iPhone in 2007. Cutting edge technology was introduced in the smartphone with an interface that allowed users to interact with the device through various modes. Also,the ecosystem facilitated developers to create applications on mobile platforms, popularly referred to as “Mobile Apps”. Mobile apps are software applications designed to run on smart-phones, tablets, and other mobile devices. They are typically available through app stores operated by entities that own the corresponding mobile operating system. By 2020, its been forecasted that mobile apps shall generate around 189 billion USD [49] in revenues via app stores and in-app advertising. The increasing use of smart-phones has led to mobile applications replacing or supplementing traditional web-based applications. Given the limitations of the form factor in smart-phones, Usability is one of the essential attributes that determine the success of a mobile apps. The measures available for assessing the usability of mobile apps tend to focus more on the human aspects and less on the functional aspects of Usability. Also, quantitative metrics that measure usability of mobile apps are very limited. As part of our initial research, we studied mobile apps in banking and finance domain. An analysis of the usability issues reported by the mobile users helped capture the challenges faced by these users as part of their day-to-day transactions. Various market giants have proposed varied usability guidelines to enrich user experience (like ISO, Android, Neilson Norman group etc). These guidelines are the repository for software practitioners to build usable mobile apps. However, there are no definitive standards or effective tools to detect and address usability issues. This made us propose a few banking specific usability guidelines based on a usability index called Mobile App Usability Index (MAUI).This study helped us establish an understanding of the functional and non-functional aspects of usability features while building a mobile app. Traditionally, usability evaluation is considered as a post-development activity which is usually taken care of by UX/UI Experts. Usability evaluation was done manually using various methods like cognitive walk-through, field studies, group studies etc, which delays detection of usability issues and requires re-development to address violations. To address this challenge, we formulated a usability framework called UMETRIX to conduct an automated usability evaluation for mobile app developers. The framework is flexible and developer friendly. The framework helps software practitioners validate and verify the presence of a usability guideline in a mobile app using code analysis. We built an automated tool using our framework that has the flexibility to create developer friendly use-cases to evaluate the presence of a guideline along with a code-snippet recommendation engine to ease developers to choose a code template for implementing the desired usability guideline. We consider this framework as a productivity tool for developers to pre-validate and evaluate usability issues during the development phases of a mobile app. As part of our research, we collaborated with open source software mobile app vendors for a field study. Additionally, we conducted industrial empirical study on a few FinTech app vendors and captured their observations and results. We collaborated with industry practitioners in their mobile app development to address mobile usability issues on their pre-released version of the mobile app. The collaboration helped us document the guidelines which are possible for evaluation using code analysis.

Abstract

Parallel algorithms on large graphs play a key role in various problems from several domains of sciences and engineering. Of these, symmetry breaking problems such as matchings and colorings are fundamental owing to a large number of applications. Algorithms for symmetry breaking problems are studied in several parallel/distributed settings over the decades. However, as the size of graphs corresponding to real-world phenomenon increase, it is imperative to use not only just parallelism but also algorithmic enhancements based on the structure of real-world graphs. A popular approach in this context is to use a graph decomposition to break the problem into multiple subproblems the solutions of which can be used to solve the original problem. A big question in this direction that is left unanswered by current research is to study which decomposition is appropriate for a given computation and a target architecture. In this context, we address the above question with respect to four problems: Maximal Matching (MM), Vertex Coloring (COLOR), Balanced Coloring (BCOL), and Approximate Weighted Matching (AWM). For these four problems, we study four different decomposition techniques including one based on bridges, one based on a random partitioning, one based on vertices of degree k for a given k, and the popular METIS based decomposition. We show that existing algorithms for the above computations on a multi-core CPU and a GPU can be significantly improved by making use of an appropriate decompo- sition of the input graph. Our study indicates that the exact decomposition to use depends both on the problem and the target architecture.

Abstract

Text Classification is among the primary challenges of natural language processing with multitude of applications including but not limited to sentiment analysis, emoji prediction and topic modeling. Several approaches ranging from rule-based systems to machine learning models have been applied to solve text classification. Neural network models, especially, have shown promising results. However, the models are limited by their reliance on the amount of annotated data. Machine learning models, primarily, operate on numerical data that represent the features of its input. In the case of text classification, the models need extraction of text features relevant to the given problem. This requirement led to the advent of a more specific area analyzing feature engineering called representation learning. Current approaches to learning sentence representations has predominantly been centered around the semantics of a sentence. However, semantic features are often inconsequential in text classification problems. Thus, to bolster task-specific representation learning, models need to learn text features according to the problem, rather than adopting a pre-compiled representation model. With the proliferation of Internet and its penetration into multilingual societies, the linguistic diversity of the real world is now reflected in online communities. Limiting solutions to a set of major languages is no longer viable. Nevertheless, the proliferation is relatively recent and hence the amount of available data in many widely-spoken languages is inadequate. Manual annotation of data is a humongous task and often expensive. Additionally, morphologically rich languages’ property of inflection and agglutination causes data sparsity problems. The possibility of utilizing resource-rich languages as leverage to enhance the performance of text classification in resource-poor languages is fascinating. To this end, the thesis presents a twin Bidirectional Long Short Term Memory (Bi-LSTM) network with shared parameters consolidated by a contrastive loss function (based on a similarity metric). Fundamentally, the aim is to classify text into multiple categories based on its features. The model jointly learns the text features (or representations) of resource-poor and resource-rich languages in a mutually shared space by utilizing the similarity between their assigned categories. Shared parameters of the Siamese network enable the projection of sentences into a common space and the similarity metric ensures the correctness of the projection. Essentially, the model projects sentences with similar categories closer to each other and sentences with different categories farther from each other. Additionally, this enables the leveraging of resource-rich languages to enhance text classification of resource-poor languages. The reason is that the projection is based on the language-agnostic annotation tags. So, pairing resource-poor and resource-rich sentences together as input to the Bi-LSTMs pair will result in their projection into a shared space. Hence, the resource-rich sentences aid the classification of resource-poor sentences by composing more training samples. Experiments, specifically, on the text classification tasks of Multilingual Sentiment Analysis and Multilingual Emoji Prediction with validation against large scale standard datasets reveal that the model significantly outperforms the state-of-the-art approaches in the case of both resource-poor and resource-rich languages. Furthermore, it is also observed that jointly training resource-poor and resource-rich languages exhibit significant performance enhancement over its single resource-poor language counterparts.

Abstract

Aesthetics is the study of the science behind the concept and perception of beauty. Psychologists have long been studying it via various experiments, which is called as experimental aesthetics. Aesthetic attributes are a set of universal principles to create aesthetically pleasing art objects, based on human’s visual processing. These attributes are taught as basic design principles, for example visual balance, color harmony, etc. In this thesis, we investigate the perception of visual balance in abstract black and white and colored paintings with experiments with university students. Visual balance is defined as a certain arrangement of the elements in a picture with respect to the canvas and with respect to each other, the point at which the composition appears more stable and coherent. We derive an objective measurement of visual balance based on the center of visual weight. The visual weight of an element is the attention-pull that element exerts on the viewer. We later synthesize an approach for optimal text placement over images using the same objective measurement and validate the results with user studies. With advancement in computer vision and Machine Learning, there have been many attempts to model the aesthetics especially in paintings, and photographs [42, 23, 71]. In this thesis, we also model eight different aesthetic attributes via a multi-task deep convolution network and derive state of art results on AADB dataset [28]. We also visualize these aesthetic attributes via gradient based technique and discuss the different level of complexity associated with different attributes.

Abstract

Enhanced Sampling in Molecular Dynamics simulations is one of the main challenges of computational chemistry. The problem arises due to existence of local minima in the free energy landscape, which serve as strong attractors. The problem is tackled by various traditional methods by somehow biasing the walker out of the minima to explore rest of conformational space. In this work we propose a multiple replica method that introduces repulsive interactions between them so as to diverge them in conformational space. These trajectories however follow a time dependent hamiltonian, statistics of which can only we studied in the nonequilibrium regime. We use the landmark work of Jarzynski to put together the statistics from these walkers to reconstruct the equilibrium free energy landscape. We also describe an implementation of the proposed method in the popular MD software, PLUMED so as to facilitate easy usability of the method for the larger community.

Abstract

Modern cryptography is based on the availability of Public Key Infrastructure (PKI) and Digital Signatures. However these schemes are designed around the assumed computational hardness of some problems like the integer factorisation, discrete logarithm, and lattice-based problems among others. This hardness assumption may turn out incorrect in future. So it is not advisable to design and implement real-world systems based on an unproven assumption. Also with the decrease in hardware costs and with the advent of new computing paradigms such as the quantum computing, the security of quite a few of such techniques is proven to be ineffective. For instance, one can potentially break RSA by implementing Shore’s Algorithm on a quantum computer. Hence, we must consider information-theoretically secure schemes which are based on the worst case assumption that the adversary has unbounded computational power. In solving problems in the design and implementation of fault-tolerant distributed systems, a basic construct that is assumed to be present is that every pair of nodes in the communication network has a channel between them for communication. This communication channel enables direct delivery of messages between them in a reliable and secure manner without the help of other nodes. However, due to the large scale of real-world distributed systems, a direct channel between every pair of nodes is not feasible/economical. Therefore, nodes depend on other intermediary nodes for communication out of which some can be faulty. We consider the setting where a subset of nodes can be under the control of a computationally unbounded adversary. The problem of simulating a reliable and secure channel from a special node S called the sender to another special node R called the receiver is popularly known as ”Secure Message Transmission (SMT)”. In this thesis, we focus on the variant of SMT protocols called Perfectly Secure Message Transmission (PSMT). The goal of PSMT is to design a protocol that enables S to communicate a secret message m to R assuring the following two conditions: (1) perfect reliability, that is, by the end of the protocol R gets the correct message m without any error and (2) perfect secrecy i.e., the adversary cannot learn any information about m, whatsoever, in information theoretic sense. It is well-known in the literature that the feasibility of PSMT protocols is influenced by the timing model. We distinguish between two variants to timing models: Synchronous and Asynchronous. While the synchronous model guarantees that the message sent by a node gets delivered to another node after a fixed amount of time, in an asynchronous model there is no such upper bound. Meaning, message delivery from one node to another node can be arbitrarily delayed. While assuming the network to be completely synchronous is too optimistic an assumption to make, assuming the network to be completely asynchronous makes solutions hard. In this work, we establish the concept of serendipitous synchrony with a focus on its influence/utility on the feasibility of PSMT protocols in asynchronous networks. We ask the following natural question: ”If part of a given asynchronous network by fortune turns out to behave synchronously (serendipitous synchrony), then does it affect the feasibility of PSMT protocols?”. We take the similar approach taken in literature and model the network as a set of disjoint wires, say n in number, with nodes in-between behaving as mere routers, popularly known as wires model. We assume the most powerful adversary considered in the literature viz., Byzantine adversary which can control at most t of these n wires arbitrarily. We quantitatively model serendipitous synchrony in the network as the number of wires that fortunately, happen to be synchronous, say ns (ns < n). Turns out that contrary to our intuition that serendipitous synchrony improves fault-tolerance of an asynchronous network, we discovered that, using serendipitous synchrony we either gain all the fault tolerance or we gain nothing from it! We further improve our insights into asynchronous networks by studying them in the presence of failstop faults (the nodes which can crash at will at any point during the execution). The main challenge in this setting is the inability to distinguish between a slow sender and a faulty sender which may sometimes cause otherwise correct protocols to have non-terminating executions. Unlike the approach that is taken in literature, where the network is abstracted as an undirected graph consisting of a set of n node disjoint wires, where each wire from S to R is also a wire from R to S; we consider the network with its directionality. That is, a wire from S to R may not be a wire from R to S. We consider mixed adversary that can cause dual failures. That is adversary can control at most tp wires in passive fashion (it can eavesdrop on messages on these wires) and at most tf wires in fail-

Abstract

The concentration of atmospheric CO 2 and antibiotic drug effluents from industries continue to increase beyond safety threshold levels in the environment. Mitigation of these contami- nants is necessary to protect the environment from various fatal consequences including global warming and antibiotic resistance caused by them. Expandable clay minerals are emerging as promising candidates for long-term storage of various environmental pollutants and nuclear wastes. The layered aluminosilicate clay minerals permit intercalation of water molecules, simple and complex organic and biomolecules into their interlayer galleries. The intercalated molecular species in the interlayer space cause the clay to swell. The swelling nature of clay is exploited in recent years for storage of pollutants including CO 2 and radioactive materials. The anthropogenic CO 2 is collected from industries and injected into deep geological formations such as saline aquifers and oil reservoirs beneath the earth’s surface at supercritical conditions (323 K, 90 bar). This process is known as geological sequestration of CO 2 . These natural geological formations are sealed on top by cap rocks that ensure leak-proof storage of CO 2 . Cap rocks are made up of swelling clay minerals including montmorillonite and bei- dellite. The CO 2 retention capacity and adsorption/desorption energetics of layered materials are likely to depend critically on the hydration level and the nature of molecular interactions among H 2 O, CO 2 , charge-balancing cations, and the oxide/hydroxide layers. The porosity and permeability of clay minerals are also greatly influenced by intercalated molecules and exchangeable charge balancing cations (CBCs). Thus, molecular-scale understanding of the structure, dynamics, and interfacial energetics of H 2 O/CO 2 binary mixtures confined in the interlayer nanopores is paramount to geological CO 2 storage efforts in clay-rich materials. Antibiotics are commonly used to treat a large variety of infections. These antibiotic drugs are produced in massive scales in India and China. Effluents from antibiotic drug in- dustries in many countries including India are directly released into nearby streams without proper effluent treatment strategies. Recent studies reported that the antibiotic drug con- centration in some streams in India and China is ∼ 100,000-1 million times higher than the prescribed threshold drug levels in the surface water. These drugs ultimately settle in clay 8 layers and enter the human body through food and water causing health hazards including an- tibiotic resistance among people residing around these streams. Due to the dynamic disorder exhibited by intercalates, the atomistic details of physico-chemical properties of these drugs in clay interlayers are not readily amenable to experimental techniques. Molecular dynamics simulations could provide atomistic level insights into the structure, dynamics, and energetics of clay-confined drug molecules. The microbial degradation of plant and animal tissues in the soil gives rise to complex heterogenous soft matter known as the natural organic matter (NOM). NOM plays a crucial role in many geochemical processes in the soil including sorption of toxic metal ions, bio- chemical regulators, radionuclides, organic contaminants, and nutrient transport and they also act as reservoirs of carbon in the soil. Despite significant research efforts in this area of re- search, the molecular structure of NOM is not fully understood because of the fact that the composition of NOM depends on the vegetation, climate, and topography of specific ecosys- tem. Recent experiments describe NOM as heterogeneous supramolecular mixtures of rela- tively small molecules, which are held together by weak non-covalent interactions between various chemical moieties of NOM. NOM-clay interactions play an important role in regu- lating the presence of carbon content in the soil. Much remains to be understood about the atomic-level details of the behavior of NOM at clay-water interfaces. In this thesis, we have employed molecular dynamics simulations and enhanced sam- pling free energy methods to elucidate the atomistic details of the structure and dynamics of H 2 O/CO 2 , antibiotics, and NOM at clay interfaces. In Chapter 3 & 4, we have studied the structure, dynamics, and energetics of H 2 O/CO 2 in smectite clay interlayers. The role of charge balancing cations in the adsorption of CO 2 at mineral-water interfaces is discussed in Chapter 5. In Chapter 6, we have derived force field parameters for antibiotic drug molecules and modeled drug-clay composites using newly generated force field parameters. The inter- actions of different functional groups of NOM, their structure, orientation and affinity towards cations, in clay-water interlayers are discussed in Chapter 7.

Abstract

Quantum information theory has gained a lot of attention recently due to its non-intuitiveness and its promise to help in developing new technologies like quantum cryptography, "quantum computer" which are known theoretically to handle very complex and intractable problems. A bit is the fundamental unit of classical information theory. But in quantum world there is qubit, which can be in both the states simultaneously termed as superposition. Superposition is a fundamental brick in quantum information theory. Due to it many types of cool phenomenons like quantum correlations i.e. entanglement, coherence etc. come into play and act as quantum resources for certain information theoretic tasks. It has been proved that quantum resources for completion of such tasks are more optimized and efficient than their classical counterparts. There are many restrictions imposed on quantum world in terms of no-go theorems out of which one we are interested in is "No-cloning" theorem. It states that there does not exists any quantum operation which can clone an arbitrary quantum state perfectly. Quantum Cryptography is one of the use case of such restrictions as it makes it more secure. We have proposed even more stronger restriction which is termed as "No-cloning of quantum coherence/superposition". Why stronger you may ask? It is known that cloning of arbitrary quantum state implies signaling. So the reverse is that no-signaling implies no-cloning of quantum state. Similarly, cloning of quantum state implies cloning of quantum coherence. So the reverse is that no-cloning of quantum coherence implies no-cloning of quantum state. This states that no-cloning of quantum coherence is stronger that no-cloning of quantum state. Also we have proved how they are different from each other as one might think at first that they are the same. The theorem holds true with or without machine states in the picture. We then show state characterization i.e. given a approximate coherence cloner how many states are there whose coherence can be cloned perfectly. Another important thing of this paper is that the theorems are proved geometrically with a nice representation on the Bloch sphere. The state of a quantum system is described by a wave function. But in cloning of a quantum state, we try to clone both wave and particle nature of the system. While in cloning of quantum coherence, we just try to clone the wave aspect of the system. From this we can also comment on how it is related to Heisenberg’s uncertainty principle. Coming to the another part of this thesis work. Another fascinating thing of quantum information theory is that we can use quantum resources to outperform some classical tasks or to perform some tasks whose classical counterparts does not exist. Quantum Superposition and Quantum Entanglement are used as quantum resources in various information theoretic tasks like Quantum ix teleportation, Super dense coding, entanglement swapping, remote state preparation, quantum cryptographic protocols, digital signature etc. It begins with the age old uncertainty principle which arises from non-commutativity of mutually complementary operators, on one face forbids us from doing many things whereas on the other face becomes the seed of great advantage in the domain of security and privacy.The fidelity of quantum information processing tasks like quantum teleportation, dense coding and entanglement swapping are related to Fully Entangled Fraction (FEF) which in turn also serves as an index that characterize non-local correlations and gives us a measure how much far a state is from the maximally entangled state. Here we try to approximately broadcast the FEF of an arbitrary quantum state acting as a resource which in turn broadcasts the fidelity of various information processing tasks. This process is extremely helpful when we need more number of quantum states with good resource abilities. We also prove no-broadcasting of non-locality using a very famous B-H state dependent local cloner. So to conclude, this thesis is all about a few impossibilities and approximations in quantum information theory.