Abstract

Planar region understanding from a single RGB image constitutes two essential components – depth map estimation and surface normal map estimation. In this work, we focus on the latter. We have proposed an approach to predict surface normals for each pixel. While surface normal prediction problem has been extensively studied, most approaches are geared towards indoor scenes and often rely on multiple modalities (depth, multiple views) for accurate estimation of normal maps. Outdoor scenes pose a greater challenge as they exhibit significant lighting variation, often contain occluders, and structures like building facades are often ridden with numerous windows and protrusions. Conventional supervised learning schemes excel in indoor scenes, but do not exhibit competitive performance when trained and deployed in outdoor environments. Furthermore, they involve complex network architectures and require many more trainable parameters. To tackle these challenges, we present an adversarial learning scheme that regularizes the output normal maps from a neural network to appear more realistic, by using a small number of precisely annotated examples. We target surface normal estimation in outdoor scenes – for building facades, using only RGB image input. Our method presents a lightweight, simpler, end-to-end trainable, single view architecture that utilizes skip connection, residual learning and adversarial regularization to generate fairly accurate normal maps of buildings in the city scenes while improving performance by at least 1.5x across most metrics. We provide qualitative and quantitative comparisons of our method against the state-of-the-art methods on normal map estimation on two city scene datasets - one real, and another synthetic, and observe significant performance enhancements. These evaluations demonstrate how even without using other modalities like depth, our single-view method is able to outperform state-of-the-art methods.

Abstract

Personal Assistants like Amazon’s Alexa, Google Home, and Apple Siri, have become a part and parcel of our everyday life. These assistants not only work towards answering user queries but also aim to emulate a human conversation by generating interactive responses for user queries. In this thesis, we look at a two-fold manner of the addition of interactive nature in such open-domain chatbots. We observe that topic diversion occurs frequently with engaging open-domain dialogue systems like virtual assistants. The balance between staying on topic and rectifying the topic drift is important for a good collaborative system. To resolve this we work towards segmenting the conversation in a flat hierarchical manner with novel annotation tags of a major topic, minor topic, and off-topic. Here, we present a model which uses a fine-tuned XLNet-base to classify the utterances pertaining to the major topic of conversation and those which are not, with a precision of 84%. We propose a preliminary study, classifying utterances into their respective annotation tags, which further extends into a system initiative for diversion rectification. This task of classifying utterances into those which belong to the major theme or not would also help us in the identification of relevant sentences for tasks like dialogue summarization and information extraction from conversations. In addition, a case study was conducted where a system initiative is emulated as a response to the user going off-topic, mimicking a common occurrence of the mixed-initiative present in the natural human-human conversation. An important feature of our topic-shift detection system is that it doesn’t o perate with any pre-defined topic sets, and detects the shift in a dynamic manner, as expected in interactions between two humans. In addition, we also look at the requirement of automatic question generation, which works towards increasing the interactive nature and conversational ability of a personal assistant. This thesis presents a system that automatically generates multiple, natural language questions using relative pronouns and relative adverbs from complex English sentences. Our system is syntax-based, runs on dependency parses information of a single-sentence input, and achieves high accuracy in terms of syntactic correctness, semantic adequacy, fluency, and uniqueness. One of the key advantages of our system, in comparison with other rule-based approaches, is that we nearly eliminate the chances of getting a wrong wh-word in the generated question, by fetching the requisite wh-word from the input sentence itself. Depending upon the input, we generate both factoid and descriptive type questions. Our contribution towards the exploitation of wh-pronouns and wh-adverbs to generate questions is novel in the Automatic Question Generation task. We explore in this thesis, the intersection between these two fields, which helps us in not only generating interactive responses but also helping the user by guiding them towards their goal.

Abstract

Leptoquarks (LQs) are hypothetical particles that connect the quark and the lepton sectors of the Standard Model (SM). As a result, the phenomenology of leptoquarks becomes very interesting and has been extensively studied in recent years. However, the majority of the recent attention is mainly due to their ability to explain the long-standing anomalies observed in B-meson decays and the muon anomalous magnetic moment measurements. Generally, the models explaining the anomalies contain LQs that couple with the third-generation quarks and leptons. This gives additional motivation to search for such LQs (commonly known as the third-generation LQs in the literature) at the Large Hadron Collider (LHC). In this thesis, we explore a novel search strategy to efficiently identify third-generation scalar leptoquarks at 14 TeV High Luminosity (HL)-LHC with an integrated luminosity of 3 ab−1 . LQs here decay to a top quark plus a tau lepton (tτ) and carry a charge of either ±1/3 (φ1) or ±5/3 (φ5). Our analysis utilizes a combined signal which consists of events from both single and pair production channels of LQ to improve the discovery prospects. We have divided our search into different benchmark scenarios based on LQ branching fractions and lepton-quark coupling parameters. We make use of several opensource HEP packages to simulate the experiments. A Monte Carlo event generator called Madgraph is used to generate the signal and background processes at the leading order (LO) with higher-order corrections being applied to some of the processes. Parton showering and hadronization is performed using the Pythia package while the detector effects are simulated using Delphes. Once all the processes are generated, we introduce a series of event selection cuts that are optimized to retain events from both pair and single production processes while rejecting a majority of background events. Finally, we try to estimate the discovery reach and exclusion mass limits of scalar LQs projected at the HC-LHC integrated luminosity of 3 ab−1 . Analysing the final result, we conclude that the overall HL-LHC reach goes beyond 1.09 (1.33) TeV for φ1 with BR = 0.5 (1.0) and 1.33 TeV for φ5 with BR = (1.0). In the first chapter, We begin our discussion with a brief introduction to the Standard Model (SM), we then talk about some of its limitations and mention a few beyond the SM (BSM) theories. The next chapter provides an overview of the experimental particle setup at the LHC along with a brief summary of jet reconstruction algorithms and jet grooming techniques. Chapter three mainly talks about the LQ phenomenology and model details while providing a sneak peek of a recent LQ search performed at the CMS. In the fourth chapter, we discuss different aspects of our analysis including the simulation setup and search strategy. Finally, we present the results and conclusions in the fifth chapter.

Abstract

The use of the Internet and online social networks has increased tremendously around the world over the past decade, providing an opportunity to exchange thoughts, ideas, opinions, and feelings with other people. This exponential growth of social media networks has allowed the production, distribution, and consumption of data at a phenomenal rate. However, it has led to various forms of online problems such as offensive content, hate speech, fake news, racism, trolling, etc. Additionally, such issues arise due to the ability to express any content on social media and the possibility of writing anonymous posts or comments. Although every social media platform has community standards and guidelines against online abuse, often these guidelines are indistinct and subjective. So, it is vital to identify and mitigate such social media issues to prevent psychological impact on affected communities and to avert hate crimes. And simultaneously, we need to increase encouraging, positive, and supportive content on social media. This thesis predominantly focuses on developing automated models to detect and classify the (abovementioned) online issues in two distinct modalities. The first one is Text; another one is the combination of Text + Vision (multimodality). In the text modality, we worked on two different types of text, one of which is the monolingual English text and the other is code-mixed text. Code-mixing is a prevalent phenomenon that occurs in text or speech on social media, so we work on social media problems with the monolingual and code-mixed text. Internet memes appear in multimodality (text + vision) on social media networks. Their primary purpose is to communicate ideas and opinions through the use of combinations of text and images, which create a specific state for the receptor, depending on the message the meme is to send. Most of the memes are downright humorous, while others, behind an amusing presentation, seek to convey subtle subtleties, including hatred, fake, sarcasm, propaganda regarding an idea, or a motivational purpose. Therefore, it is necessary to identify and remove such content on social media and make it a safer place for everyone. Hence, this thesis presents the various traditional machine learning models to the latest advanced transformer-based models to detect burgeoning social media problems in text and multimodalities. Coming to language use in social media, we find both monolingual data and code-mixed data. Initially, in monolingual text, we investigate the work of detecting fake news in COVID-19 related social media posts through the adaption of transformers-based contextual word representations. We propose an ensemble model by concatenating the BERT, ALBERT, and XLNet representations. This contextual word representations-based ensemble model outperformed all other individual transformer models for this problem. Group of people increasing conflict and hatred among consumers by exploiting freedom of speech and expression on social media platforms. Therefore, it is essential to take a positive reinforcement approach to research on positive, helpful, and supportive social media content. Wherefore, we build a transformer-based BERT with a threshold-based language detection system to detect Hope Speech in Youtube comments. The primary intent of this work is to reduce negativity and strengthen encouraging, supportive, and efficacious social media content.Next, we examine two principal issues in code-mixed data. Language identification is a primary preprocessing step in numerous code-mixed applications. With this purpose, we develop a word level language identification system by utilizing the English-Telugu code-mixed content. Subsequently, we explore the problem of sentiment analysis with English-Hindi bilingual code-mixed data. For this task, we propose a character level and sub-word level word representations with long-short term memory architecture. Later in multimodality, we explore a very salient new issue on social media: the emotion analysis of Internet memes. This work analyzes three different types of issues that revolve around Internet memes. The first of them is predicting the sentiment polarity of a meme, the second is a multi-label classification task that assesses whether a meme is offensive, humorous, satirical, or motivational, and the third is a multi-output ordinal classification task that predicts the degree of offense, humor, sarcasm of a meme. To handle these tasks, we introduce a multimodal architecture with late fusion technique that combines LSTM for textual features with VGG-16 for image feature representations. Propaganda is a communication tool that nfluences the opinions and actions of other people to achieve a predetermined goal. Initially, it was seen in newspapers, advertisements, and so on, but now it is widely used on social media. Therefore we develop a multimodal fusion system to detect propaganda in memes. We have used a robust fusion

Abstract

Evapotranspiration (ET) is one of the prominent hydrologic variables affecting water and energy balances and critical factors for crop water requirements and irrigation scheduling. ET is a complex hydrological variable defined by various hydro climatological variables. The major forms of ET which are widely applicable in hydrological, water balance, drought assessments and other ecological assessments are Potential (PET) and Actual Evapotranspiration (AET). PET represents the atmospheric water demand with a focus on climatological variables. AET is influenced by climate, vegetation, soil moisture, and the amount of available water and presents the annual water balance between precipitation and latent heat exchange. Various empirical formulations have been developed to estimate PET and AET depending upon the availability of hydrometeorological variables. These empirical formulations are region-specific and developed for particular climatic conditions. Furthermore, the empirical family of models has a major limitation as they require a large number of hydrometeorological inputs, limiting their utility in data-scarce areas of ungauged basins. In this context, empirical and mathematical models have emerged as simple and readily implementable for estimating PET and AET with measured hydrometeorological parameters as independent variables. Such mathematical models can be valuable to predict PET and AET when climate data is insufficient. The present study compared various empirical models and data-driven algorithms to predict PET and AET using various hydroclimatic variables. Four empirical methods, such as FAO-based Penman-Monteith method, temperature-based Hargreaves method, and radiation-based Turc, and Priestley-Taylor method, were used to estimate PET at a daily time scale. Five data-driven algorithms, such as Long short-term memory neural networks (LSTM), Artificial Neural Network (ANN), Gradient Boosting Regressor (GBR), Random Forest (RF) and Support Vector Regression (SVR), were implemented. Two empirical AET models, such as Budyko and Turc methods, were utilized in estimating AET. These models were evaluated over two different climatic regions, Hyderabad, the largest city of the Indian state, Telangana and aipara in New Zealand, both with semiarid climates. Dataset consists of daily meteorological data of maximum and minimum air temperature, relative humidity, solar radiation, and wind speed over a period of 51 years (1965 - 2015) for Hyderabad and for a period of 6 years (2010- 2016) for Waipara station.The Penman-Monteith method was considered as the standard method to compare the different models and various empirical models of PET. The models were trained and tested with climate variables as input variables and various empirical models as reference models. The most influencing climate variables on PET were found in the order of temperature, solar radiation, wind speed, and relative humidity, which formed as the basis to choose different datasets to train over models and compare the results to validate. Temperature and radiation-based models of Turc and Priestley-Taylor methods can be used to estimate PET when all other climate variables are not available as they are also promising with the Penman-Monteith method. The results indicated that 99% accuracy could be achieved with all climatic input, whereas accuracy drops to 86% with limited data. Both LSTM and ANN models have been noted as the most robust models for estimating PET with minimal climate data. Even though the excellent performance can be achievable when all input variables are used, the study, however, found that even a three-parameter combination (temperature, wind speed and relative humidity values) or two-parameter combination (temperature and relative humidity, temperature and wind speed) can also be promising in PET estimation for a semi-arid climate. The AET over semi-arid climatic conditions of Hyderabad, Telangana, India and Waipara (New Zealand) was estimated using modelled and different empirical methods-based PET using Budyko and Turc models. The proposed empirical-based AET models, Budyko and Turc, showed that the AET process has the potential to be estimated by structurally simple methods. Equation-based AET methods made it possible to extract useful information about the hydrological process. It was observed that the meteorological variables of temperature and solar radiation have more significant contributions than other variables in the estimation of AET. In addition, the effects of the meteorological variables were found to be essential and effective in the estimation of AET. The research findings of the study reveal that under limited data availability, the best input combinations were identified as temperature and wind speed for estimating PET; temperature, wind speed and precipitation for estimating AET for semiarid climatology. Overall, the research findings of the study stress on the use of limited data in

Abstract

Quantum computing hardware has progressed rapidly. Simultaneously, there has been a proliferation of programming languages and program optimization tools for quantum computing. Existing quantum compilers use intermediate representations (IRs) where quantum programs are described as circuits. Such IRs fail to leverage existing work on compiler optimizations. In such IRs, it is non-trivial to statically check for physical constraints such as the no-cloning theorem, which states that qubits cannot be copied. We introduce QSSA, a novel quantum IR based on static single assignment (SSA) that enables decades of research in compiler optimizations to be applied to quantum compilation.QSSA models quantum operations as being side-effect-free. The inputs and outputs of the operation are in one-to-one correspondence; qubits cannot be created or destroyed. As a result, our IR supports a static analysis pass that verifies no-cloning at compile-time. The quantum circuit is fully encoded within the def-use chain of the IR, allowing us to leverage existing optimization passes on SSA representations such as redundancy elimination and dead-code elimination. Running our QSSA-based compiler on the QASMBench and IBM Quantum Challenge datasets, we show that our optimizations perform comparably to IBM’s Qiskit quantum compiler infrastructure. QSSA allows us to represent, analyze, and transform quantum programs using the robust theory of SSA representations, bringing quantum compilation into the realm of well-understood theory and practice.

Abstract

The goal of rendering in computer graphics is the synthesis of photorealistic images from abstract virtual scene descriptions. Among the many rendering methods, path tracing along with physically accurate material modeling has emerged as the industry standard for photorealistic image synthesis. Path tracing is, however a computationally expensive operation. Moreover, physically accurate material modeling requires iterative parameter tuning and visualization using path tracing. This quickly becomes a bottleneck for the tuning of a large number of parameters found in most physically accurate material models, and artists end up spending a lot of time waiting for path tracing to converge. This dissertation proposes to leverage the learning power of neural networks to plausibly approximate path tracing for: (a) quick material visualization while editing, (b) out-of-the box material recovery using inverse rendering. The traditional workflow for tuning material parameters and visualization is typically followed by modifications to lighting for better understanding the behaviour of the material. We thus incorporate the ability to modify the environment lighting in our neural material visualization framework. This significantly aids in understanding the behaviour of the material thus aiding the process of parameter tuning, which we demonstrate with a user study. Most real-world materials are spatially varying, which is that their behaviour varies across the surface of the geometry. While previous works in this area dealt with uniform materials, we propose improvements for handling spatially varying materials. We design our neural network architecture to be light-weight, having 10× lesser trainable parameters than the state-of-the-art, while also producing better visualizations. We also build a tool to demonstrate real-time parameter tuning and visualization of materials. Finally, we extend the above work to handle general 3D scenes in specific lighting scenarios. This is done by training a small fully connected neural network on a simple scene in the local tangent frame. Training in the local tangent frame decouples geometric complexity from training which implies that the network generalizes to any geometry. We embed this network in a standard path tracer to evaluate direct illumination at each path vertex. We thus not only obtain noise free direct illumination but also out-of-the box gradients of the rendering process, which can be used for inverse rendering. We show that these gradients can be used for plausibly recovering the material of a target scene. We conclude by discussing various practical aspects of our method.

Abstract

Recent advances in networking sites, the Internet revolution has changed everything, right from businesses to healthcare to education to the ways of communicating with our friends. The Internet has opened doors for people to express themselves, write their thoughts about a particular topic, share an experience online for other people to read without much hassle. Even before going out, order something, or watching a show/movie, people tend to check the online reviews first. And it is feasible and practical also because anyone wouldn’t want to spend their time, money, and other resources on something that isn’t worth it. With so many reviews, tweets, and content available online, it is also important to process them in such a way that they can be used by everyone productively. One use-case could be as the word limit for online reviews on sites like IMDb, Zomato, and Amazon is pretty significant (10, 000 characters for IMDb, Amazon has a limit of 5, 000 words ∼ 23, 500 characters) some reviews tend to be longer. While the reviewer elaborates on their experience, from a reader point of view, what really important is, what aspects of the given target entities the reviewer liked/disliked. This is where the need for Aspect-based Sentiment Analysis arises. Aspect-Based Sentiment Analysis (ABSA) aims to extract the aspects of the given target entities and their respective sentiments. The main issue is, the amount of data is so huge that manually processing the data on such a vast scale is impossible. E.g., Twitter alone sees an average of 6000 tweets per second, roughly 500 million tweets per day. Hence, we look for some fast, automated methods that can do the processing almost in real-time. In this thesis, we propose a deep-learning enabled model, TEASER, based on an extract-then-classify framework for extracting the aspects and detecting the respective sentiment attached. We also conduct extensive experiments on the 3 existing datasets (Restaurant14, Restaurant15, Laptop14) to show that TEASER performs better than the existing models. In chapter 4, we present two novel datasets in the domain of movie reviews, Movie20, and moviesLarge. Movie20 is a supervised dataset of 1162 sentences manually annotated by two human annotators, whereas moviesLarge is a pseudo-labeled dataset of 14373 sentences. With the help of Semi-supervised learning, we benchmark TEASER on the Movie20 dataset. We also evaluate the results of TEASER on the Movie20 dataset thoroughly and try to reason the gaps in predicted output and gold annotation.

Abstract

Pattern mining is an important task of data mining and involves the extraction of interesting associations from large databases. In the literature, pattern mining approaches have been investigated to extract the different kinds of patterns such as frequent patterns, sequential patterns, periodic patterns, utility patterns and coverage patterns from transactional databases. The pattern mining approaches have been widely employed to improve the performance of applications in the areas of recommendation systems, e-commerce, customer relation management, internet advertising, decision support systems, social media and so on. One of the key issues of developing pattern mining algorithms is as follows: a pattern mining algorithm has to process a huge volume of transactional dataset. Research efforts are going on to develop pattern mining algorithms to operate on large datasets by exploiting different parallel processing paradigms, namely shared memory, shared disk, shared-nothing, GPUs, incremental, MapReduce and Grid. Typically, a given transactional database D gets updated due to the addition and/or deletion of transactions. Consequently, some of the previously discovered patterns may become invalid, while some new patterns may emerge. To extract the new knowledge we should mine the database every time it gets updated. One of the main issues is to mine huge Dynamic Databases for every time instant from scratch is computationally costly. The concept of incremental mining is evolved to solve the issue of mining patterns for every instant efficiently for Dynamic Databases and has become an active research area. When the database is updated, with Incremental mining, there is an opportunity to improve the performance by limiting the computation to updated parts of the database and with the required additional processing. In the literature, Incremental mining approaches have been investigated to extract different kinds of patterns such as frequent patterns, sequential patterns, periodic patterns and utility patterns from transactional databases. In this thesis, we propose an improved approach for extracting Coverage Patterns for Dynamic Databases under the incremental paradigm. Given a transactional database, the coverage pattern mining involves the extraction of multiple sets of items, where each set (covers a certain percentage of transactions) satisfies relative frequency, coverage support and the overlap ratio constraints. The applicability of coverage patterns has been demonstrated in improving the performance of search engine advertising, banner advertising and visibility mining. In the literature, a level-wise iterative coverage pattern mining (CPM) and a pattern growth approach have been proposed.In this thesis, we propose a Comprehensive Coverage Pattern Mining (CCPM) approach, for efficiently extracting Coverage Patterns under the incremental paradigm, when a set of transactions are added only, set of transactions are deleted only and both set of transactions are added and deleted from the original database. For each case, we have identified a set of rules by identifying the potential opportunities to avoid unnecessary processing as compared to the rudimentary option of the processing from the scratch. Furthermore, it has been observed that, when a set of transactions are added and/or deleted from the original database, if the order of items in the frequency list (Flist) of the database is changed, additional patterns have to be validated and require additional processing. Based on the rules and by considering the issue of change in the order of items in the Flist, we have proposed improved approaches to extract coverage patterns, by considering the case of (i) addition of transactions (ii) deletion of transactions and (iii) addition and deletion of transactions. We have also performed extensive experiments on two real click-stream datasets and one synthetic dataset and demonstrated the overall effectiveness of our proposed approach.

Abstract

The recent advances in Machine Learning and other data driven techniques have transformed domains like robotics, computer vision and natural language processing. These algorithms are now also being applied to a wide array of chemistry problems as well. One such class of machine learning algorithms is deep generative modelling which has found its application in de novo molecule generation. Generating lead molecules in silico, helps reduce the wastage of limited computational and laboratory resources by reducing the number of generated bad leads. For every task, the lead molecules may exist in different regions of the chemical space which would have to be explored appropriately. The drug discovery process is long, arduous and expensive while also being plagued by the problems caused by bad leads especially for novel targets and the design of new inhibitors for novel targets is now more important than ever especially in the current scenario with the world being plagued by COVID19. Drug molecules must possess certain properties to be effective so that they can be considered for further validation. Conventional computational approaches such as high-throughput virtual screening require extensive combing through existing database and evaluating all the molecules in them in the hope of finding possible hits that possess the required properties. However, existing libraries of druglike molecules, advances in deep generating modelling and optimization techniques like reinforcement learning has provided a new avenue to tackle this problem. In this thesis, a computational strategy is proposed for the de novo generation of molecules with high binding affinities to the specified target and other desirable properties for drug-like molecules using reinforcement learning. A deep generative model built using a stack-augmented recurrent neural network initially trained to generate drug-like molecules is optimized using reinforcement learning to start generating molecules with desirable properties like LogP, Quantitative Estimate of Drug Likeliness,Topological Polar Surface Area, and Hydration Free Energy along with the binding affinity. For multiobjective optimization, a novel strategy was also devised in which the property being used to calculate the reward is changed periodically. In comparison to the conventional approach of taking a weighted sum of all rewards, this strategy shows an enhanced ability to generate a significantly higher number of molecules with desirable properties while also exploring new regions of the chemical space for multiple targets.