Abstract

Social media sites such as Twitter and Facebook have emerged as popular tools for people to express their opinions and sentiments on various topics. Large amount of data provided by these media is extremely valuable for mining trending topics and events. However, this massive volume also dictates that the mining approach be efficient in terms of computations and storage. In this thesis, we propose an efficient, scalable system to detect events from tweets. The system does not employ any Twitter-specific features, and thus, can be readily adapted to any other social media site. Since tweets are very short and diverse in nature, traditional approaches, meant for typical long/well-formatted documents, can not be applied to tweets. Tweets are written informally with a lot of abbreviations and mistakes. Since tweets are very short, statistical concepts like tf.idf can not be directly used on them. Our approach detects events by exploring their textual and temporal components. The system does not require any target entity to be specified; it automatically detects generic events from a set of tweets. The key components of our system are an extraction scheme for event representative keywords, an efficient storage mechanism to store their appearance patterns, and a hierarchical clustering technique based on the common co-occurring features of keywords. We evaluate our system on two data-sets, one provided by VAST challenge 2011 and other published by US based users in January 2013. Our approach is efficient in terms of computational time and memory, and achieves high precision across two different datasets. The detected events are easy to interpret, and encompass a wide range of topics.

Abstract

Sponsored search is one of the most dominant modes of online advertising on the web. In sponsored search, advertisers bid on relevant keywords to advertise their product. For an incoming search query, advertisements from the ad campaigns containing the query keywords are shown along with the search results. If multiple advertisers demand to be shown on the same query’s results page, they are ranked for the allocation of ad space. The ranking is determined by multiple factors including the bid amount of the advertiser on the query keywords, relevance of ad content to the search query, Click-Through-Rate (CTR) and budget of the advertiser. The ecosystem of sponsored search has three main stakeholders - search engine, advertisers and users. Search engines aim to maximize the revenue by showing ads. Advertisers want to maximize the reach of their service or increase the sales. Users see advertisements on the search results page. Some of the key research challenges in sponsored search are query-ad relevance, click-through-rate prediction, optimal auction design, optimum utilization of ad space of rare (tail) queries and click-fraud detection. In this thesis, related to sponsored search, we have investigated approaches to exploit the ad space of tail queries. It is well established that search queries tend to follow a heavy-tailed Zipf distribution wherein a large fraction of queries occur too infrequently. Such infrequent query set is called long tail. Advertising on long tail queries is challenging as long tail queries are encountered rarely which makes them hard to interpret for sponsored search. Also, it has been observed that during keyword auctions, advertisers tend to bid for the head keywords to reach more users. This creates a high demand for the head query keywords and little or no demand for the tail query keywords. The long tail phenomenon also makes it quite difficult for an advertiser to capture the relevant keywords from the long tail. The above stated factors result in under-utilization of a significant amount of the ad space provided by tail queries in sponsored search which is identified as the research issue. We propose two approaches for the utilization of ad space of tail queries by proposing that instead of bidding on keywords, advertisers should bid upon high level concepts. In the first approach, we propose bidding on concepts by organizing concepts into a two level taxonomy. In the second approach, we propose a generalized approach by considering organization of concepts into a multi-level taxonomy. In the first approach, we have proposed an improved framework to cover more advertisers by exploiting the notions of coverage and concept taxonomy based on the log data of search queries. The proposed framework allows to form the distinct groups of keywords such that each group of keywords could be allocated to meet the demands of the advertiser. We model each search session as a transactionand queries occurring in a session form the items of the transaction. Coverage patterns are then mined from these session-based transactions. The coverage patterns fuse tail keywords together into multiple groups to maximize the unique visitors and two-level concept taxonomy ensures that groups are meaningful. A comprehensive framework has been proposed to map the extracted coverage patterns and the demands of the advertisers to allocate incoming queries to advertisers. We have conducted experiments on a real world dataset of AOL web search engine and found out that it is possible to meet the demands of more advertisers with the proposed approach. It was found out that the proposed approach is able to provide diverse but meaningful group of keywords which allows the advertiser to display advertisement to appropriate users based on the requirements. In the second approach, we propose that advertisers should bid upon high level concepts represented by a multi-level taxonomy instead of search keywords during ad space auctions. Advertisers are free to bid on any node of the taxonomy. Bidding on any node in the taxonomy provides more flexibility to the advertisers to target the potential consumers. However, to allocate children nodes of bidding nodes to advertisers, the flat model of coverage patterns cannot be used. To address the issues of interdependency of concepts on each other, we exploit search query logs and a taxonomy to extract level-wise coverage patterns. We further propose an end-to-end architecture which takes search query logs, axonomy and advertising demands as inputs and allocates an incoming search query to the advertisers. The corresponding architecture is used to perform allocation of incoming queries to advertisers for sponsored search. Experiments on a real world dataset of AOL search query logs show improvement in performance with respect to ad space utilization and reach of the advertisements. Overall, we have proposed two approaches for improving the utilization of ad

Abstract

Autonomous Navigation in indoor environment where human and robot occupy the same workspace is a challenging problem, robots should behave in a socially acceptable way otherwise they will cause discomfort for humans which could possibly lead to confusion and disastrous results. In recent years autonomous navigation systems have been a primary focus in robotics. A wheelchair is a mechanical device that can often assist people with disabilities. It comes under Rehabilitation Technologies or assistive technologies. It effectively uses wheels and mechanical support to overcome a loss of legs or leg control. Manual wheelchairs can be operated by persons who can use their upper body or someone available to assist. They can help people who have motor control problems, spinal injuries and amputation. An autonomous wheel chair can also help people who have lost motor control of their legs as well as their arms. They can help in moving patients in hospitals as well as moving heavy equipments in a warehouse like environment. The thesis contributes in the following ways. Firstly, we develop a wheel chair capable of autonomously navigating in an indoor environment. We also develop a novel approach to predict human intentions and introduce a probabilistic variant of time scaled collision cone constraints. These constraints allows us to relate space of collision free velocities to an associated confidence measure. Secondly, it develops an optimization framework for extracting a specific solution from the entire solution space of collision free velocities, depending on the factors such as distance from goal, probability of collision , robots kinematics and dynamic constraints such as turning radius, maximum acceleration and maximum velocity.

Abstract

In recent years, there has been a surge in finding vulnerabilities in browsers. This is mainly attributed to the ability of browser to host various services by gluing different applications. However, most of the research has gone into addressing web-related vulnerabilities like, XSS, SSL, SQLi etc. In this work, we focus on other type of vulnerabilities namely low-level or implementation level bugs. We specifically target embedded interpreters in browsers, as these software process codes that may come from a malicious site. Our contribution is to advance state of the art in interpreter testing. Fuzzing is an automated black box testing technique used for finding security vulnerabilities in the software by providing random data as input. Several frameworks for typical applications (e.g. network protocols, files formats, API‘s, applications and other targets) have been written so far. However, for interpreter testing, very few fuzzers exist. We present an automated evolutionary fuzzing technique to find bugs in JavaScript interpreters. However, in the case of an interpreter, fuzzing is challenging because the inputs are piece of codes that should be syntactically / semantically valid to pass the interpreter’s elementary checks. On the other hand, the fuzzed input should also be uncommon enough to trigger exceptional behavior in the interpreter, such as crashes, memory leaks and failing assertions. In our approach, we use evolutionary computing techniques, specifically genetic programming, to guide the fuzzer in generating uncommon input code fragments that may trigger exceptional behavior in the interpreter. We implement a prototype named IFuzzer to evaluate our technique on real-world examples. IFuzzer uses the language grammar to generate valid inputs. We applied IFuzzer first on an older version of the JavaScript interpreter of Mozilla (to allow for a fair comparison to existing work) and found 40 bugs, of which 12 were exploitable. On subsequently targeting the latest builds of the interpreter, IFuzzer found 17 bugs, of which four were security bugs.

Abstract

The thesis presents end-to-end frameworks for joint semantic and motion segmentation in outdoor dynamic scenes in both monocular and stereo setup. Dynamic scene understanding is a challenging problem and motion segmentation plays a crucial role in solving it. Incorporating semantics and motion enhances the overall perception of the dynamic scene. For applications of outdoor robotic navigation, joint learning methods have not been extensively used for extracting spatio-temporal features or adding different priors into the formulation. The task becomes even more challenging without stereo information being incorporated. The thesis proposes an approach to fuse semantic features and motion clues using CNNs, to address the problem of monocular semantic motion segmentation. We deduce semantic and motion labels by integrating optical flow as a constraint with semantic features into dilated convolution network. The pipeline consists of three main stages i.e Feature extraction, Feature amplification and Multi Scale Context Aggregation to fuse the semantics and flow features. Our joint formulation shows significant improvements in monocular motion segmentation over the state of the art methods on challenging KITTI tracking dataset. The focus in the second part of the thesis is temporally consistent joint semantic and motion segmentation. Segmenting moving objects in a video sequence has been a challenging problem and critical to outdoor robotic navigation. While recent literature have laid focus on regularizing object labels over a sequence of frames, exploiting the spatio-temporal features for motion segmentation has been scarce. Particularly in real world dynamic scenes, existing approaches tend to fail in segmenting moving objects with large camera motion. In this thesis, we present an approach for exploiting semantic information and temporal constraints in a joint framework for motion segmentation in a video. We propose a formulation for inferring per-frame joint semantic and motion labels using semantic potentials from dilated CNN framework and motion potentials from depth and geometric constraints. We integrate the potentials obtained into a 3D(space-time) fully connected CRF framework with overlapping/connected blocks. We solve for a feature space embedding in the spatio-temporal space by enforcing temporal constraints using optical flow and long term tracks as a least-squares problem. We evaluate our approach on KITTI Tracking benchmark and demonstrate results superior to the state-of-the-art in motion segmentation.

Abstract

Natural language processing is a field of artificial intelligence and computational linguistics that aims at bridging the gap between human beings and computers. It deals with processing natural languages in various forms like speech and text. Processing a natural language happens at various levels i.e, word, phrase, sentence, syntax, semantics and discourse. In this work we present our efforts at making advancements at syntax and semantic levels in dependency parsing. Dependency parsing involves discovering relationships between words in a given sentence so as to understand it better. We begin this thesis with a comparative error analysis of two parsers - MALT and MST on Telugu Dependency Treebank data. We discuss the performance of both the parsers in relation to the Telugu language and then talk in detail about both the algorithmic issues of the parsers as well as the language specific constraints of Telugu. We follow this by proposing a semi-supervised approach by introducing clustered word embedding based features into dependency parsing. This involves a simple and effective semi-supervised method to introduce features that incorporate dependency label clusters derived from a large annotated corpus into data driven dependency parsing. We demonstrate the effectiveness of the approach in a series of experiments on Hindi language data. We then propose a neural network based classifier voting approach to dependency parsing using multiple classifiers as component systems in an ensemble and a neural network algorithm as an oracle. We show significant improvements over the best component systems for both transition-based and graph-based dependency parsing.

Abstract

SiliconCarbide(SiC) is a wide band gap semiconductor with high thermal conductivity, high electron saturation velocity and high critical electric field properties. Silicon based devices, if replaced by SiC devices, could reduce the switching losses and improve the power conversion efficiency in grids becauseof their high frequency capability. However, the main limitation of SiC based MOS devices is thei r inability in handling higher electric fields at the oxide interface. An alternative to overcome this limitation is to replace the conventionally used SiO 2 with the high-K dielectrics that can sustain high electric fields. High-K dielectrics ZrO 2 and HfO 2 , with a dielectric constant of about 25, have been chosen for this study, as SiO 2 , with a dielectric constant of 3.9, cannot sustain high electric fields. This approach has been attempted earlier with HfO 2 as the gate dielectric in the case of SiC power MOSFETs. Previous studies of SiC MOSFETs with HfO 2 /SiO 2 stacking gate dielectric have shown significant improvement in the electronic performance. The electric field in the high-K dielectrics does not exceed the safe operational value, even when the underneath SiC has exceeds high electric field values. In order to address these concerns, an asymmetric trench gate SiC IGBT using high-K dielectrics such as HfO 2 and ZrO 2 as alternatives to the conventionally used SiO 2 has been studied. While supporting high electric field is one aspect, in this work it is investigated whether replacement of SiO 2 with the high-K dielectrics considered has any influence on the switching characteristics of IGBTs such as rise time, fall time, and turn OFF time using TCAD (Technology Computer-Aided Design) simulations. The static and switching characteristics of the SiC IGBT have been explored using technology based two-dimensional numerical computer simulations. ZrO 2 and HfO 2 exhibited better forward transcon-ductance ratio and reduced tail current thereby indicating improved power losses. The power dissipation curves were found to be consistent with the results obtained. However, these high-K dielectrics have low band offset as compared to SiO 2 -SiC interface. A lower band offset could lead to a high gate leakage current, which was resolved by introducing a thin layer of SiO 2 between SiC and high-K dielectrics. Power dissipation has been analyzed for IGBT with SiO 2 , HfO 2 and ZrO 2 layers by applying a pulse signal. In addition, the power dissipation is simulated for temperature range of 300-600K for different dielectric layers as the switching performance of a device is supposed to deteriorate with increase in temperature. HfO 2 and ZrO 2 demonstrated good temperature stability. The work in this paper is an attempt to correlate the power dissipation with the static and dynamic characteristics for the respective oxide layers.

Abstract

A brain template, such as MNI152 is a digital (magentic resonance image or MRI) representation of the brain in a reference coordinate system for the neuroscience research. Structural atlases, such as AAL and DKA, delineate the brain into cortical and subcortical structures which are used in Voxel Based Morphometry (VBM) and fMRI analysis. Many population specific templates, i.e. Chinese, Korean, etc., have been constructed recently. It was observed that there are morphological differences between the average brain of the eastern and the western population. In this thesis, we report on the development of a population specific brain template for the young Indian population. This is derived from a multi-centeric MRI dataset of 100 Indian adults (21 - 30 years old). Measurements made with this template indicated that the Indian brain, on average, is smaller in height and width compared to the Caucasian and the Chinese brain. A second problem this thesis examines is automated segmentation of cortical and non-cortical human brain structures, using multiple structural atlases. This has been hitherto approached using computationally expensive non-rigid registration followed by label fusion. We propose an alternative approach for this using a Convolutional Neural Network (CNN) which classifies a voxel into one of many structures. Evaluation of the proposed method on various datasets showed that the mean Dice coefficient varied from 0.844±0.031 to 0.743±0.019 for datasets with the least (32) and the most (134) number of labels, respectively. These figures are marginally better or on par with those obtained with the current state of the art methods on nearly all datasets, at a reduced computational time. We also propose an end-to-end trainable Fully Convolutional Neural Network (FCNN) architecture called the M-net, for segmenting deep (human) brain structures. A novel scheme is used to learn to combine and represent 3D context information of a given slice in a 2D slice. Consequently, the M-net utilizes only 2D convolution though it operates on 3D data. Experiment results show that the M-net outperforms other state-of-the-art model-based segmentation methods in terms of dice coefficient and is at least 3 times faster than them.

Abstract

Trivia is any fact about an entity which is interesting due to its unusualness, uniqueness or unexpectedness. Trivia could be successfully employed to promote user engagement in various product experiences featuring the given entity. A Knowledge Graph (KG) is a semantic network which encodes various facts about entities and their relationships. We propose a novel approach called DBpedia Trivia Miner (DTM) to automatically mine trivia for entities of a given domain in KGs. The essence of DTM lies in learning an Interestingness Model (IM), for a given domain, from human annotated training data provided in the form of interesting facts from the KG. The IM thus learnt is applied to extract trivia for other entities of the same domain in the KG. We propose two different approaches for learning the IM - a) A Convolutional Neural Network (CNN) based approach and b) Fusion Based CNN (F-CNN) approach which combines both hand-crafted and CNN features. Experiments across two different domains - Bollywood Actors and Music Artists reveal that CNN automatically learns features which are relevant to the task and shows competitive performance relative to hand-crafted feature based baselines whereas F-CNN significantly improves the performance over the baseline approaches which use hand-crafted features alone. We also study the problem of relevance scoring of triples. We use the triples with “type-like” relations in which we employed a dataset with relevance scores ranging from 0 to 7, with 7 being the “most relevant” and 0 being the “least relevant”. The task focuses on two such relations: profession and nationality. We built a system which could automatically predict the relevance scores for unseen/new triples. Our model is primarily a supervised machine learning based one in which we use well-designed features which are used to a build a Logistic Ordinal Regression based classification model.

Abstract

Slow moving traffic in heavily populated cities, can many times result in loss of lives due to emergency vehicles not being able to reach their destination hospitals on time. Recent advances in the field of Intelligent Transportation Systems (ITS) makes it increasingly likely that vehicles in the near future will be equipped with advanced systems that allow inter vehicular communication. In this thesis, we assume the usage of such a system to optimize the lane level dynamics for an emergency vehicle (EV), traversing a multi lane stretch of road under a variety of traffic settings. In particular, we present the Fixed Lane Strategy (FLS) and the Best Lane Strategy (BLS) for EV traversal and perform an extensive agent based analysis to study their strengths and weaknesses. Through a series of experiments performed using the well-known traffic simulator SUMO, we could show that: (a) BLS performs better than SUMO strategy on all traffic settings we tested. (b) BLS performs better than FLS in settings that capture real-world traffic conditions involving congestion and uncertainties while FLS performs better in well-behaved conditions and (c) BLS was found to be the best strategy for the setting calibrated using real world data (obtained from NYCDOT).