Abstract

Social dynamics is based on human needs for trust, support, resource sharing, irrespective of whether they operate in real life or in a virtual setting. Massively multiplayer online role-playing games (MMORPGS) serve as enablers of leisurely social activity and are important tools for social interactions. Past research has shown that socially dense gaming environments like MMORPGs can be used to study important social phenomena, which may operate in real life, too. We describe the process of social exploration to entail the following components 1) finding the balance between personal and social time; 2) making a choice between a large number of weak ties or few strong social ties; 3) finding a social group. In general, these are the major determinants of an individual’s social life. This research looks into the phenomenon of social exploration in an activity-based online social environment. We study this process through the lens of the following research questions, 1) What are the different social behavior types? 2) Is there a change in a player’s social behavior over time? 3) Are certain social behaviors more stable than others? 4) Can longitudinal research of player behavior help shed light on the social dynamics and processes in the network? We use an unsupervised machine learning approach to come up with 4 different social behavior types - Lone Wolf, Pack Wolf of Small Pack, Pack Wolf of a Large Pack and Social Butterfly. The types represent the degree of socialization of players in the game. Our research reveals that social behaviors change with time. While lone wolf and pack wolf of small pack are more stable social behaviors, pack wolf of large pack and social butterflies are more transient.We also observe that players progressively move from large groups with weak social ties to settle in small groups with stronger ties. In addition to the above, this research also analyzes the churn behavior of different social types. Churn broadly refers to the loss of users from a system or network. The churn behavior of different social behavior types was analyzed to see if there is a link between social behavior and churn activity. The analysis reveals that churn propensity of different social behavior types is different. Behaviors characterized by a higher degree of socialization have a significantly lower propensity to churn than behavior characterized by a lower degree of socialization. We studied the social trends by looking at the data from a popular MMORPG and aimed to look into the dynamics and social phenomena of the network through the lens of different social behaviors.

Abstract

The ever-growing demand for livestock product because of increase in population, urbanisation and income create a need for development in this sector. Advancements in nutrients and creature well-being will keep on adding to expanding potential generation and further proficiency. Sorghum is important forage crop used in the development of livestock. It is equipped for delivering significant amounts of feed amid the hotter months of the year. However, the livestock production not only relies on the accessibility of grain but the continuous availability of good quality forage. Assessment of quality of a forage depends on the estimation of nutrients present in it. There is a need for improvement in the time-consuming and exorbitant traditional methods for measuring nutrients in sorghum fodder. Near InfraRed Spectroscopy (NIRS) is the new and growing technology in this field. It is the quickest method for estimating nutrient, yet the high expense of the spectroscopic instrument limits its utilization. Therefore, there is a necessity to explore techniques which are low cost, inexpensive and are not labourintensive. NIRS technology scans the spectra for a wide range of wavelengths, which makes it expensive. We have applied Partial Least Square Regression (PLSR) on NIRS data of Sorghum to select significant wavelengths for Crude-Protein (CP), Neutral Detergent Fiber (NDF) and Acid Detergent Fiber (ADF) content estimation. Calibration models proposed using a single wavelength (1449nm) and a pair of wavelength (1485nm and 1486nm) give the best predictive squared correlation coefficient (Q2 ) as 0.7928 and 0.8334 respectively for CP. A pair of wavelengths (1165 nm and 2391 nm) is giving the maximum Q2 value as 0.4045 for NDF. For ADF, the maximum value of Q2 obtained is 0.2208 using 354 nm and 1204 nm wavelengths are used in combination. An LED based chemical detection technique has been used to calculate absorbance for CP estimation at one of its important wavelengths. The advantages of LEDs are that they are low cost, low power and readily available important optical device. The wavelength of 1450 nm giving the Q2 value as it is a commercial LED which is easily available in the market and has been used to design hardware in combination with a Photo Diode (PD) of same peak wavelength to measure the absorbance of collected samples of sorghum. The absorbance values, obtained by using the above mentioned hardware design, are then used in a calibration model equation of 1450 nm wavelength, derived using NIRS dataset by applying PLSR. A laboratory analysis method for CP calculation known as the Kjeldahl method is applied to the same samples for verification purposes. On comparing the CP value obtained using LED based method with that of the Kjeldahl method, we are getting a deviation of 19.48% Through this work, we can implement the fastest, low cost and precise design for the estimation of CP values for sorghum. Further, we can incorporate this design for other nutrient values such as NDF and ADF using their respective set of significant wavelengths.

Abstract

Extensive research is being carried out in the field of Natural Language Processing (NLP) in the context of Indian languages. Spelling correction, word segmentation, and grammar checking are the fundamental problems in NLP. The aim of these problems is to identify noise in the data and correct it. These tools are important for many NLP applications like web search engines, text summarization, sentiment analysis, machine translation etc. Many methods have been developed for these problems for English, which usually exploit linguistic resources like parsers, large amounts of real world data etc. making it difficult to adapt them to other languages. For these problems, deep learning models have also been implemented for English. These models use parallel data of noisy and correct mappings from different sources as training data for automatic correction tasks. Indian languages are resource-scarce and do not have such parallel data due to low volume of queries and non-existence of such prior implementations. Spell correction is very crucial for applications like web search engines. In speech recognition and optical character recognition (OCR), word boundaries are not often comprehensible. Grammar checking is important to clean up the corpus before using it in any application. It is also useful as an aid for second language users and for applications like automatic essay grading. In applications like speech-to-text, handwritten text-to-text, and Machine Translation, applying spell checking, word segmentation, and grammar checking on the output will result in improved accuracy. Most of the existing language processing systems for Indian languages are rule-based and language specific. Majority of the works have a dictionary-based approach and check the errors against the dictionary. Due to lack of availability of training data for machine learning models, we create synthetic datasets using language rules. In this work, we present approaches to build automatic language correction systems for Hindi. We have proposed deep learning models for spelling correction, word segmentation, and grammar checking. The proposed approaches are applicable to any resource-scarce language. A comparative evaluation for each of these three problems shows that the proposed models are competitive with the existing correction techniques for Indian languages.

Abstract

Semantic role labeling, also known as shallow semantic parsing is a task in Natural Language Processing to determine labels of the words or the phrases (group of words) in a sentence. The labels are of the type agent, patient/receiver, goal, temporal, locative or object. It comes under the domain of Artificial Intelligence for the reason that we want to do this task automatically. Formally, the task can be broken into various steps. The first step is to identify the verbs or the predicates in a sentence. The second step consists of detection of the semantic arguments associated with the corresponding predicates in the sentence and the final step would be to label these arguments pertaining to their respective predicates. For instance, in the sentence, “Ram killed Shyam with a gun”, the semantic role labeler should be able to recognize ‘killed’ (represents the phrase “to kill”) as a predicate. Then ‘Ram’ as representing the killer (agent), ‘Shyam’ as the recipient/receiver and “a gun” as representing the theme/object. Such representations are an important step towards figuring the meaning of a sentence. Semantic role labeling (SRL) gives a lower level of abstraction compared to a full syntactic parsing of a sentence. Therefore, it has higher number of classes and groups fewer clauses in each class. For example, “the car belongs to Ram” will need two labels like ‘owned’ and ‘owner’ whereas “the car was sold to Ram” will need two different labels altogether such as ‘theme/goal’ and ‘receiver’ even though these two clauses would be similar if we use subject and object functions. Hence, the labels largely depend on the whole clause rather than just the individual phrase or chunk. In this thesis, we begin with a detailed overview of the literature in the field of semantic role labeling. This includes discussion over the various techniques used to tackle shallow semantic parsing in the past, the development of different datasets built for semantic analysis such as the Propbank and the Framenet and at last we talk about semantic role labeling task for Indian languages. Since most of the Indian Languages are very low-resourced, we have kept our work restricted to only Hindi and Urdu, for which the respective propbanks were good in size. We also give a detailed analysis of the Hindi/Urdu Propbank data sections used by us in our experiments.Next, we present a statistical semantic role labeler for Hindi which is extended to Urdu as well. We propose a set of new features enriching the existing baseline system for these languages. We break the system into two subsequent tasks - Argument Identification and Argument Classification respectively. Our experiments show a reasonable improvement with respect to the previous baseline for Hindi, mainly for the classification step. We also report significant improvements for Argument Identification task in Urdu. We propose a new baseline system for Indian languages using 5-fold cross-validation and we capture results excluding the ‘null’ class and including the ‘null’ class exclusively. Lastly, we give an error analysis of our model compared with the previous baseline and look at both the improvements and limitations of this model. Finally, we introduce a supervised deep learning model for the same Indian languages namely, Hindi and Urdu which uses minimal syntax and yet improves over our statistical model significantly. We progress with three different models inspired from the recent advancements in this field. In the first model we make use of sequence modeling to generate dependency path embeddings and jointly learning the classification process i.e., both Identification and Labeling of arguments. The second model is a syntax agnostic model where we encode the full sentence using a bi-directional LSTM encoder only using the raw words/tokens. The third and the final model adds dependency label to the previous model making it slightly syntax-aware and performs very well compared to the other models. At last, we talk about evaluation metrics and analysis of the three models as well as the statistical model.

Abstract

We live in an increasingly connected world, where directed links in social connectivity graphs could represent anything from longstanding friendships or aligning niche interests to a publisher-subscriber relationship quenching an insatiable appetite for real-time news or celebrity gossip. A pertinent way of describing the relation between social media and its users can be summarized by rescripting cultural anthropologist Michael Wesch’s quote: “Social Media is Us/ing Us” This alludes to the phenomena where social mores and personal tastes shape the landscape of our online platforms such as popular hashtags and trending topics while these platforms simultaneously also shape the attitudes and outlook of the people who use them. Social media platforms are just as amenable to be used for proselytizing impressionable users as they are to getting long ignored or suppressed viewpoints out in the open. In the face of a runaway fake news problem and faced with a world where a superior online presence could decide elections, policy matters and foment prejudices, it is unfortunate that as a result of a myriad of factors including poor platform design and vacillating stances on the significance of platform badges by social networks, the social interaction aspects of authenticity and prominence become inextricably linked with one another and frequently conflated, the consequences of which can be drastic. It has been shown that the presence of authenticity markers next to user-generated content enhances its reach and credibility, while the posts from highly perceived sources have been demonstrated to likely be identified as more plausible. Social network and publishing platforms, such as Twitter, support the concept of verification. Ver- ified accounts are deemed worthy of platform-wide public interest and are separately authenticated by the platform itself. There have been repeated assertions by platforms a la Twitter and Facebook, about verification not being tantamount to endorsement. However, a significant body of prior work suggests that possessing a verified status symbolizes enhanced credibility in the eyes of the platform audience. As a result, such a station is highly coveted among public figures and influencers. Hence, we attempt to characterize the network of verified users on Twitter and compare the results to similar analyses performed for the entire Twitter network. We extracted the whole graph of verified users on Twitter (as of July 2018) and obtained 231,235 English user-profiles and 79,213,811 connections. Subsequently, in the network analysis, we found that the sub-graph of verified users mirrors the full Twitter users graph in some aspects, such as possessing a short diameter. However, our findings contrast with earlier results on multiple fronts, such as the possession of a power-law out-degree distribution, slight dissortativity, and a significantly higher reciprocity rate, as elucidated in the thesis. Moreover, we attempt to gauge the presence of salient components within this sub-graph and detect the absence of homophily with respect to popularity, which again is in stark contrast to the full Twitter graph. Finally, we demonstrate stationarity in the time series of verified user activity levels.It is in this backdrop that we attempt to deconstruct the extent to which Twitter’s verification policy mingles the notions of authenticity and authority. To this end, we seek to unravel the aspects of a user’s profile, which likely engender or preclude verification. The aim of the thesis is two-fold: First, we test if discerning the verification status of a handle from profile metadata and content features is feasible. Second, we unravel the characteristics which have the most significant bearing on a handle’s verification status. We augmented our dataset with all the 494 million tweets of the aforementioned users over a one year collection period along with their temporal social reach and activity characteristics. Our proposed models are able to reliably identify verification status (Area under curve AUC > 99%). We show that the number of public list memberships, presence of neutral sentiment in tweets and an authoritative language style are the most pertinent predictors of verification status. To the best of our knowledge, this work represents the first quantitative attempt at characterizing verified users on Twitter and also the first attempt at discerning and classifying verification worthy users on Twitter.

Abstract

Visual object tracking is a fundamental task in computer vision and is a key component in wide range of applications like surveillance, autonomous navigation, video analysis and editing, augmented reality etc. Given a target object with bounding box in the first frame, the goal in visual object tracking is to track the given target in the subsequent frames. Although significant progress has been made in this domain to address various challenges like occlusion, scale change etc., we observe that tracking on a large number of short sequences as done in previous benchmarks does not clearly bring out the competence or potential of a tracking algorithm. Moreover, even if a tracking algorithm works well on challenging small sequences and fails on moderately difficult long sequences, it will be of limited practical importance since many tracking applications rely on precise long-term tracking. Thus, we extend the problem of visual object tracking for long time horizons systematically in this thesis. First, we first introduce a long-term visual object tracking benchmark. We propose a novel largescale dataset, specifically tailored for long-term tracking. Our dataset consists of high resolution, densely annotated sequences, encompassing a duration of over 400 minutes (676K frames), making it more than 20 folds larger in average duration per sequence and more than 8 folds larger in terms of total covered duration, compared to existing generic datasets for visual tracking. The proposed dataset paves a way to suitably assess long term tracking performance and train better deep learning architectures (avoiding/reducing augmentation, which may not reflect real world behaviour). We also propose a novel metric for long-term tracking which captures the ability of a tracker to track consistently for long duration. We benchmark 17 state of the art trackers on our dataset and rank them according to several evaluation metrics and run time speeds. Next, we analyze the long-term tracking performance of state of the art trackers in depth. We focus on the three key aspects of long-term tracking: Re-detection, Recovery and Reliability. Specifically, we (a) test re-detection capability of the trackers in the wild by simulating virtual cuts, (b) investigate the role of chance in recovery of tracker post failure and (c) propose a novel metric allowing visual inference on the contiguous and consistent aspect of tracking. We present several insights derived from an extensive set of quantitative and qualitative experiments. Lastly, we present a novel fully convolutional anchor free siamese framework for visual object tracking. Previous works utilized anchor based region proposal networks to improve the performance of siamese correlation based trackers while maintaining real-time speed. However, we show that enumerating multiple boxes at each keypoint location in the search region is inefficient and unsuitable for the task of single object tracking, where we just need to locate one target object. Thus, we take an alternate approach by directly regressing box offsets and sizes for keypoint locations in the search region. This proposed approach, dubbed SiamReg, is fully convolutional, anchor free, lighter in weight and improves target localization. We train our framework end-to-end with Generalized IoU loss for bounding box regression and cross entropy loss for target classification. We perform several experiments on standard tracking benchmarks to demonstrate the effectiveness of our approach.

Abstract

Accurate localization of other traffic participants is a vital task in autonomous driving systems. State-of-the-art systems employ a combination of sensing modalities such as RGB cameras and Li- DARs for localizing traffic participants. However, with the advent and subsequent commercialization of autonomous driving, there has been an increased interest in monocular object localization (and re-construction) for urban driving scenarios. While there have been very interesting attempts to tackle thisproblem using a single monocular camera, most approaches assume that the ego car and the car to belocalized share the same road plane. This condition is often termed as the coplanarity assumption.In this thesis, we relax the aforementioned coplanarity assumption and demonstrate — to the best of our knowledge — the first results for monocular object localization and shape estimation on surfaces that are non-coplanar with the moving ego vehicle mounted with a monocular camera. We approximate road surfaces by local planar patches and use semantic cues from vehicles in the scene to initializea local bundle-adjustment like procedure that simultaneously estimates the 3D pose and shape of thevehicles, and the orientation of the local ground plane on which the vehicle stands. We also demonstrate that our approach transfers from synthetic to real data, without any hyper parameter-/fine-tuning. We evaluate the proposed approach on the KITTI and SYNTHIA-SF benchmarks, for a variety of roadplane configurations. The proposed approach significantly improves the state-of-the-art for monocular object localization on arbitrarily profiled/graded roads.

Abstract

Bistability is a well-known phenomenon and is a widely applied concept in electronic switches. This is due to the fact that it has exactly two stable outputs irrespective of the range of input given. Bistability along with hysteresis is proven to be very useful concepts in building a noise immune stable circuits. But hysteresis is defined as a phenomenon where effect lags behind the applied force, hence it adds delay to the system. This work explores the possibility of building a system, followed by a circuit which can possess both bistability and noise immunity without any delay or if possible an advancement in transition time compared to a hysteretic circuit or a basic inverter. This newly mentioned phenomenon is also seen in a good number of the natural process, and it is called as proteresis. Both proteresis and hysteresis are greek origin words where proto means ahead and hyste means late. The main motivation for this work is that in nature we observe that most of the things exist in pairs like high, low; Negative logic, positive logic; cause, effect; clockwise spin, anticlockwise spin; etc. And this work explores the possibility of the opposite pair of hysteresis and it’s a possible contribution in electronics. A detailed proteretic system model is built and is electronically verified by using OP-AMPs. A similar proteretic circuit is also built in CMOS technology using 16 transistors and also a possibility of CMOS one-stage hysteretic and proteretic circuit is proposed. This thesis also proposes a new type of switch which can possess the two states as proteresis and hysteresis similar to clockwise and anticlockwise spins or low and high. And a detailed system model and circuit show how we can shift between these two states. The simulation and experimental results of this switch support each other

Abstract

The research in the field of autonomous driving is making significant progress a lot of impor-tance is being placed on developing reliable motion planning frameworks that scale up to complex scenarios like urban traffic, highway driving, etc. Planning frameworks that cater to the current autonomous navigation stacks must have the capability to generate maneuvers that are real-time feasible and robust enough to accommodate the unpredictable behaviour of traffic participants. A typical reactive navigation scenario places a lot of emphasis on the computational efficiency of such planners, as the sensing and planning cycles should be in perfect synchronization. The current motion planning modules for autonomous vehicles consider large safety margins and use predefined paths for performing complex maneuvers like merging, overtaking, etc. These considerations generally result in maneuvers that are very conservative in nature and are far from being optimal. In a real urban like scenario, autonomous vehicles adopting such planning modules may cause situations like traffic disruption and unwanted deadlocks. In the proposed thesis we present a motion planning framework for autonomous vehicles to handle complex maneuvers like merging in dense traffic. At the primary level, we present an end to end trajectory optimization scheme that uses state of the art techniques to generate trajectories that are not only kinematically feasible but also very smooth to the extent of being jerk minimal. One of the other highlights here is an efficient Sequential Convex Optimization routine that has the capability to handle differential constraints which are very crucial for dynamic obstacle avoidance. To sum up, the proposed optimization scheme in this level maintains a fine balance between the spatial and temporal modifications that are carried out on a trajectory to avoid collisions consistently from start to goal. At the secondary level, we developed a risk based lane selection mechanism which forms the Lane selection layer of the proposed framework. This layer computes the likelihood of collision along each trajectory by deriving the probability distribution of the first order Collision cone constraint in closed form. One of the notable highlights here is the closed form nature of this probability distribution, which makes the computation of Collision likelihood very efficient. The trajectory that has least collision risk is then selected for performing the final maneuver. We show that the presented lane selection mechanism significantly reduces the temporal deviation causing the least disruption to the traffic flow.To demonstrate the efficacy of the presented framework we have evaluated it in several highway scenarios in simulations. Also, the framework is deployed on a real autonomous car without the lane selection layer and, its performance is evaluated on both static and dynamic environments.

Abstract

Retinal images are widely used to detect and diagnose many diseases such as Diabetic Retinopathy (DR), glaucoma, Age-related Macular Degeneration, Cystoid Macular Edema, coronary heart disease,and so on. These diseases affect vision and lead to irreversible blindness. Early image-based screening and monitoring of the patient is a solution. Imaging of retina is commonly done either through Optical Coherence Tomography (OCT) or Fundus photography. OCT captures cross-sectional information about the retinal layers in a 3D volume, whereas fundus imaging projects retinal tissues onto the 2D imaging plane. Recently smartphone camera-based fundus imaging is being explored with a relatively low-cost. Imaging retina with these technologies pose challenges due to physical properties of the light source, or quality of optics and sensors used or low and uneven light condition. In this thesis, we look at learning based approaches, namely neural network techniques to improve the quality of retinal images to aid diagnosis. The first part of this thesis aims at denoising OCT images, which are corrupted by speckle noise due to underlying coherence-based imaging technique. We propose a new method for denoising OCT images based on Convolutional Neural Network by learning common features from unpaired noisy and clean OCT images in an unsupervised, end-to-end manner. The proposed method consists of a combination of two autoencoders with shared encoder layers, which we call as Shared Encoder (SE) architecture. The SE is trained to reconstruct noisy and clean OCT images with respective autoencoders, and denoised OCT image is obtained using a cross-model prediction. The proposed method can be used for denoising OCT images with or without pathology from any scanner. The SE architecture was assessed using public datasets and found to perform better than baseline methods exhibiting a good balance of retaining anatomical integrity and speckle reduction. The second problem we focus on is the enhancement of fundus images acquired with a Smartphone camera (SC). SC image is a cost-effective solution for the assessment of retina, especially in screening. However, imaging at high magnification and low light levels results in loss of details, uneven illumination, noise particularly in the peripheral region and flash-induced artefacts. We address these problems by matching the characteristics of images from SC to those from a regular fundus camera (FC) using either unpaired or paired data. Two mapping solutions are designed using deep learning technique in an unsupervised and supervised manner. The unsupervised architecture called ResCycleGAN is based on the CycleGAN with two significant changes: A residual connection is introduced to aid learning only the correction required; A structure similarity based loss function is used to improve the clarity of anatomical structures and pathologies. This method can handle variations seen in normal and pathological images, acquired even without mydriasis, which is attractive in screening. The method produces consistently balanced results, outperforms CycleGAN both qualitatively and quantitatively, and has more pleasing results. Next, a new architecture is proposed called SupEnh, which handles noise removal using paired data. The proposed method enhances the quality of SC images along with denoising in an end-to-end, supervised manner. Obtaining paired data is challenging; however, it is feasible in fixed clinical settings or commercial product as it is required once for learning. The proposed SupEnh method based on U-net consists of an encoder and two decoders. The network simplifies the task by learning denoising and mapping to FC separately with two decoders. The method handles images with/without pathologies as well as images acquired even without mydriasis. The SupEnh was assessed using private datasets and found to performs better than U-net. The cross-validation results show method is robust to change in image quality. The enhancement using SupEnh method achieves 5% higher AUC for early stage DR detection when compared with original images.