Abstract

With the evolution of social media, the data present online now is more diverse than ever. The tremendous popularity of social media micro-blogging websites like Twitter, Facebook has provided users more freedom to post their content on the Internet without much restrictions. One effect of which is that users are free to express themselves in real time over the entire world network. This all adds a considerable amount of data to the already existing enormous internet, which attracts the need for research in this area. Hence, the need for a research community to mine and extract any useful information if possible from these social media texts. With that massive amount of data we need processing or methods in times when we want some answers to some specific question and at the same time save time and not crawl the entire Internet uselessly, like Named Entity Recognition. Apart from that we also try to extract morphological properties from the data available to us from these micro-blogging sites like aggression, sarcasm, gender, etc. One of the benefits that these social sites provide to their users is the availability to write content in their native languages as well as in code-mixed texts which brings the research community to focus on these low-resourced languages/texts. In this thesis, we analyze the problems of Named Entity Recognition (NER) and Aggression Detection in code-mixed social media texts and provide with a baseline system for these two problem statements which are not very well explored. We present a corpus of Hindi-English code-mixed tweets for Named Entity Recognition which is annotated following the BIO standard (Begin, Intermediate and Other). For the task of Aggression Detection, we used Hindi-English code-mixed dataset provided for the shared task in 1st Workshop on Trolling, Aggression and Cyberbullying (TRAC-1). Finally, we propose a supervised classification system which uses various machine learning techniques for prediction of Named Entity tags of the words in NER task and detecting Aggression associated with the text using a variety of features.

Abstract

Motivated by human visual system machine or computer vision technology has been revolutionized in last few decades and make a strong impact in wide range of applications such as object recognition, face recognition and identification etc. However, despite much encouraging advancement, there are still many fields which lack to utilize the full potential of computer vision techniques. One such field is to analyze the satellite images for geo-spatial applications. In past, building and launching the satellites in to space was expensive, and was big hurdle in acquir- ing low cost images from satellites. However, with technological innovations, the inexpensive satellites are capable of sending terabytes of images of our planet on the daily basis that can provide insights on global-scale economic, social and industrial processes. The significant applications of satellite imagery are urban planning, crop yield forecasting, mapping and change detection. The most obvious application of satellite imagery is to extract topological road network from the satellite images, as it plays an important role in planning the mobility between multiple geographical locations of interest. The extraction of road topology from the satellite images is formulated as binary segmentation problem in vision community. Despite of huge satellite imagery, the fundamental hurdle in applying computer vision algorithms based on deep learning architectures is unavailability of labeled data and causes the poor results. Another challenge in extraction of the roads from satellite imagery is visual ambiguity in identifying the roads and occlusion by various objects. This challenge causes many standard algorithms in computer vision research to perform poorly and is the major concern. In this thesis we develop deep learning based models and techniques that allows us to address the above challenges. In the first part of our work, we make an attempt to perform road segmentation with the less labeled data and existing unsupervised feature learning techniques. In particular, we use self-supervised technique to learn visual representations with an artificial supervision, followed by fine tuning of model with labeled dataset. We use semantic image in-painting as an artificial/auxiliary task for supervision. The enhancement of road segmentation is in direct relation with the features captured by model through inpainting of the erased regions in the image. To further enhance the feature learning, we propose to inpaint the difficult regions of the image and develop a novel adversarial training scheme to learn mask used for erasing the image. The proposed scheme gradually learns to erase regions, which are difficult to inpaint. Thus, this increase in difficulty level of image in-painting leads to better road segmentation. Additionally, we study the proposed approach on scene parsing and land classification in satellite images.

Abstract

GPUs with their massive Single Instruction Multiple Data (SIMD) capability have become attractive for scientific computation. The Kirchhoff-Helmholtz Transform (KHT) is a two dimensional integral commonly used in numerical reconstruction of the object from its experimentally measured diffraction pattern (called hologram). We explore the evaluation of KHT using GPU at various levels of optimisation.These optimisations are of two types: (a) algorithmic: exploiting the symmetries inherent in KHT, and (b) programmatic: optimizations that are specific to the GPU architecture like the lookup tables, scheduling of read/writes. From the numerical experiments, results show that the fully optimized version is a factor of 100 times faster than the naive GPU implementation.

Abstract

Given the huge impact of the Internet, online content along with social networks play an important role in forming or changing the opinions of people. Unlike traditional journalism where only certain news organizations can publish content, online journalism has given chance even for individuals to publish. This has its own advantages like individual empowerment, decentralization of power etc...but has given a chance to a lot of malicious entities to spread misinformation for their own benefit. It provides the mechanism to publish news events as they happen. As soon as the information is uploaded to the web, it is available around the globe. In these times where blogging is considered as journalism and thousands of websites are being built daily, it is unfortunate that the authenticity of the news on a topic can be undermined. People are more inclined to believe that the information is true if they encounter it before. As reported by many organizations in recent history, this even has an influence on major events like the outcome of elections. Hence, there is a great need for an automated intelligent system which can classify what stories people are reading to so that they can verify before believing or sharing such news stories. Such a system can be a bulwark to the dissemination of fake, false, conspiratorial and misleading news. The three specific sub-problems we dealt with in this thesis are: (1) Fine-grained classification of a news articles based on content, (2) Controversy detection based on the reactions like comments or likes and shares of an article on social networks, and (3) Stance detection in the wild using maximally informative sentences extracted from many articles discussing the same topic. As mentioned, given that the usage of online social media is increasing, many political parties and news sources are using this as a platform for spreading news. This helps us to collect more data about peoples opinions on a topic and utilize this data as the key to understanding some aspects regarding the controversy. In recent times, social media has been playing a role in how individuals process information and form political opinions. So, there has been a need to offer the user a view that differs from what they are mostly exposed to, for gaining the overall picture on a topic before coming to a conclusion or an opinion. One aspect that is the focus of these studies is how the social media user interaction can be used to predict a controversy. We investigate the topics related to US 2016 elections and present the results of our quantitative analysis that aims to classify the topic as being controversial or not. We propose a deep hierarchical attention neural architecture combining sentiment and semantic embeddings for fine-grained classification of news stories.

Abstract

The general elections of India in 2014 proved to be important for the study of political communication owing to the rise in influence of right wing ideology in national politics. Twitter was employed as a social media platform, by political leaders, to improve their public image and perception conveying their campaign goals to their massive number of followers. The most strategic use of Twitter was credited to Prime Ministerial candidate Narendra Modi. He was portrayed as a visionary leader who wanted to work for economic development, social empowerment, and good governance. Within this narrative,women’s empowerment debates have also undergone a multitude of changes.Through the thesis, we try to establish the nature of interactions between discussions on feminism, political communication, and technology. We study the discursive trajectory of women’s empowerment against the backdrop of the influence of right wing Hindutva ideology of India on Twitter. We employ a rigorous qualitative analysis using Thematic Study and Feminist Relational Discourse Analysis to study two important cases of women’s movements since 2014: Instant Triple Talaq, and Sabarimala Verdict. By analyzing tweets surrounding these incidents, we show a rhetorical inconsistency of the right wing leaders and supporters. We study the implications of this inconsistency on suppression of voices demanding empowerment and amplification of narratives on religion and tradition. Finally we conclude the implications of amplification and suppression of voices on emergence of the phenomenon of “controlled empowerment” of women.

Abstract

In this thesis, we present a novel approach to the Multi-object tracking (MOT) problem. Specifically, the focus is on tracking multiple traffic participants in road scenes from a monocular camera mounted on an autonomous car. While this is an extremely well-studied problem, state of the art approaches rely exclusively on appearance information (derived from image pixels) to associate bounding box detections over image sequences (aka tracking-by-detection). They fail to leverage the abundant world-space information that can be harnessed by usage of simple projective geometry, and other semantic cues. We show that, moving beyond pixels and incorporating geometric and object shape cues improves the performance of tracking-by-detection framework by leaps and bounds. We propose several 2D (appearance/pixel space) and 3D (world-space) cues and constructs pair wise costs that can readily be incorporated into any tracking-by-detection workflow. Moreover, all the proposed cues (both 2D and 3D) are computed using image sequences from a monocular camera. Although reasoning about the 3D shape and pose of object in a monocular setting is ill posed, we demonstrate that incorporating prior knowledge about the scene in form of :prior known camera height (which is the case in autonomous driving systems) , the assumption that vehicles being co-planar with the ego vehicle and with knowledge of how 3D shapes project in the image; can help in reasoning about the reverse process : the 3D object shape, pose and motion in a monocular setting. This knowledge about the scene and the objects helps in forming association costs in 3D. We train a Convolutional Neural Network to reason about semantic keypoints of vehicles, which not only provides observations to reason about objects in 3D, but provides complimentary 2D cues about the object appearance for the task of data association. The proposed costs are agnostic to the data association method and can be incorporated into any optimization framework. These costs are easy to implement, can be computed in real-time, and complement each other to account for possible errors in a tracking-by-detection framework. We perform an extensive analysis of the designed costs and empirically demonstrate consistent improvement over the state-of-the-art under varying conditions that employ a range of object detectors and exhibit a variety in camera and object motions. We showcase that using the simplest of associations frameworks (two-frame Hungarian assignment), the presented approach surpasses the state-of-the-art in multi-object-tracking on road scenes.

Abstract

One of the major challenges for information theorists is to build such communication schemes which allow participants to communicate efficiently . Index Coding , which was first introduced by Birk and Kol in 1998 [1] as coding on Demand by an Informed Source is a communication scheme dealing with a broadcast channel in which receivers have prior side information about the messages to be broadcast. Exploiting the knowledge about the side information, the sender may significantly reduce the number of required transmissions, in turn improve the efficiency of the communication over this type of broadcast channels. The index coding problem provides a simple yet rich model for several important engineering problems in network communication, such as content broadcasting, peer-to-peer communication, distributed caching, device-to-device relaying, interference management and opportunistic wireless networks. It also has close relationships to network coding, distributed storage, and guessing games. The Index Coding problem is to identify the minimum number of transmissions (optimal length) to be made so that all receivers can decode their wanted messages using the transmitted symbols and their respective prior information and also the codes with optimal length. Rank Invariant extensions of index coding problems are discussed in this thesis. Any index coding problem can be defined using an associated matrix known as a fitting matrix. An index coding problem Iext is called an extension of another index coding problem I if the fitting matrix of I is a submatrix of the fitting matrix of Iext. We first present a straightforward m-order extension Iext of an index coding problem I for which an index code is obtained by concatenating m copies of an index code of I. The length of the codes is the same for both I and Iext, and if the index code for I has optimal length then so does the extended code for Iext. More generally, an extended index coding problem of I having the same optimal length as I is said to be a rank invariant extension of I. We then focus on 2-order rank-invariant extensions of I, and present constructions of such extensions based on involutory permutation matrices. The second contribution in this thesis is obtaining optimal index codes for a specific subclass of index coding problems. It is known that the single unicast index coding problem can be represented using a side-information graph G (with number of vertices n equal to the number of source symbols). The optimal rate of index coding is known to be lower bounded by the size of maximum acyclic induced subgraph of the side-information graph, denoted by MAIS. For index coding problems with MAIS =n − 1 and MAIS = n − 2, prior work has shown that binary (over F2) linear index codes achieve the MAIS lower bound for the broadcast rate and thus are optimal. In this thesis, we use the the duality relationship between network coding and index coding to show that for a class of index coding problems with MAIS = n − 3, binary linear index codes achieve the MAIS lower bound on the broadcast rate. In contrast, it is known that there exists index coding problems with MAIS = n − 3 and optimal broadcast rate strictly greater than MAIS. Relevant figures has been included for clear understating of reader.

Abstract

High-level synthesis (HLS) has received significant attention in recent years for improving programmability of FPGAs. One could raise the level of abstraction further by using domain-specific languages (DSLs), improving productivity and performance simultaneously. PolyMage is a domainspecific language and compiler for image processing pipelines. Its PolyMage-HLS backend translates an input expressed as a DAG of image processing stages through the DSL into an equivalent circuit that can be synthesized on FPGAs, while leveraging an HLS suite. The power and area savings while performing arithmetic operations on fixed-point data type are well known to be significant over using floating-point data type. PolyMage-HLS stores data at each stage of a pipeline using a fixed-point data type (α, β) where α and β denote the number of integral and fractional bits. The integral bitwidth (α) requirement at a pipeline stage can be inferred from its range. In this work, we first propose an interval-arithmetic based range analysis algorithm to estimate the number of bits required to store the integral part of the data at each stage of an image processing pipeline. The analysis algorithm uses the homogeneity of pixel signals at each stage to cluster them and perform a combined range analysis. Secondly, we propose a software architecture for easily deploying any kind of interval/affine arithmetic based range analyses in the DSL compiler. Thirdly, we show that interval/affine arithmetic based techniques fail to take into account correlated computations across stages and hence could lead to poor range estimates. These errors in range estimates accumulate across stages, especially for iterative programs, such as Horn-Schunck Optical Flow, resulting in estimates nearly unusable in practice. Then, we propose a new range analysis technique using Satisfiability Modulo Theory (SMT) solvers, and show that the range estimates obtained through it are very close to the lower bounds obtained through profile-driven analysis. Finally, for estimating fractional bitwidth (β) requirement at each stage of the pipeline, we propose a simple and practical heuristic search algorithm, which makes very few profile passes, as opposed to techniques such as simulated annealing used in prior work. The analysis algorithm attempts to minimize the number of fractional bits required at each stage while preserving an application-specific quality metric. We evaluated our bitwidth analysis algorithms on four image processing benchmarks listed in the order of increasing complexity: Unsharp Mask, Down-Up Sampling, Harris Corner Detection and Horn-Schunck Optical Flow. The performance metrics considered are quality, power and area. For example, on Optical Flow, the interval analysis based approach showed an 1.4× and 1.14× improvement on area and power metrics over floatingpoint representation respectively; whereas the SMT solver based approach showed 2.49× and 1.58× improvement on area and power metrics when compared to interval analysis.

Abstract

One of the major challenges for information theorists is to build such communication schemes which allow participants to communicate efficiently . Index Coding , which was first introduced by Birk and Kol in 1998 [1] as Coding on Demand by an Informed Source is a communication scheme dealing with a broadcast channel in which receivers have prior side information about the messages to be broadcast. Exploiting the knowledge about the side information, the sender may significantly reduce the number of required transmissions , in turn improve the efficiency of the communication over this type of broadcast channels. The index coding problem provides a simple yet rich model for several important engineering problems in network communication, such as content broadcasting, peerto-peer communication, distributed caching, device-to-device relaying, interference management and oppoturnistic wireless networks. It also has close relationships to network coding, distributed storage and guessing games. The Index Coding problem is to identify the minimum number of transmissions (optimal length) to be made so that all receivers can decode their wanted messages using the transmitted symbols and their respective prior information and also the codes with optimal length. Optimal Index codes can be constructed using graph theoretic methods, algebraic methods, matrix completion, linear programming, interference alignment, etc. It is known that the index coding problem can be represented using a side-information graph G (with number of vertices n equal to the number of source symbols). The size of the maximum acyclic induced subgraph of side information graph, denoted by MAIS is a lower bound on the broadcast rate. For index coding problems with MAIS = n − 1 and MAIS = n − 2, prior work has shown that binary (over F2) linear index codes achieve the MAIS lower bound for the broadcast rate and thus are optimal. In this thesis, we use the the duality relationship between network coding and index coding to show that for a class of index coding problems with MAIS = n − 3, binary linear index codes achieve the MAIS lower bound on the broadcast rate. In contrast, it is known that there exists index coding problems with MAIS = n−3 and optimal broadcast rate strictly greater than MAIS. Relevant figures has been included for clear understating of reader.

Abstract

The problem of source coding is one of the central areas of information theory and coding. Designing source codes for the broadcast applications, where the receivers may have some prior information about the source, before it was sent is the key idea in this thesis. The challenge is to come with encoding schemes that exploit the side-information from the receivers in order to reduce the length of the code. This problem of source coding with side-information is called as index coding, first introduced by Birk and Kol [1] as coding on demand by an informed source. The index coding problem provides a simple yet rich model for several important engineering problems in network communication, such as content broadcasting, peer-to-peer communication, distributed caching, device-to-device relaying, interference management and opportunistic wireless networks. It also has close relationships to network coding, distributed storage, and guessing games.The Index Coding problem is to identify the minimum number of transmissions (optimal length) to be made so that all receivers can decode their wanted messages using the transmitted symbols and their respective prior information and also the codes with optimal length. Optimal Index codes can be constructed using different methods which include graph theoretic methods, algebraic methods, matrix completion, linear programming, interference alignment,etc. An index coding problem can be represented using a demand graph (with number of edges equal to the demands at receivers and number of vertices n equal to the number of source symbols). Various graph theoretic bounds were well studied in the literature. Uniprior index coding problem is studied in this thesis. For single uniprior index coding problem , in which each receiver has a single unique side-information symbol, a polynomial complexity construction of an optimal index code is known from prior work. In this thesis we model the uniprior index coding problem as a supergraph G S , and focus on a class of uniprior problems defined on special supergraphs known as generalized cycles G gc , in which the sizes of demand set and the side-information set are equal at every receiver. For such problems, we prove upper and lower bounds on the optimal broadcast rate. Using a connection with Eulerian directed graphs, we also show that the upper and lower bounds are equal for a subclass of uniprior problems. We show the NP-hardness of finding the lower bound for uniprior problems on generalized cycles, hence contrasting such uniprior problems with single uniprior problems. Finally we look at a simple extension of the generalized cycle uniprior class for which we give bounds on the optimal rate and show an explicit scheme which achieves the upper bound.