Abstract

The use of local detectors and descriptors in typical computer vision pipelines work well until variations in viewpoint and appearance change become extreme. Past research in this area has typically focused on one of two approaches to this challenge: the use of projections into spaces more suitable for feature matching under extreme viewpoint changes, and attempting to learn features that are inherently more robust to viewpoint change. In this paper we present a novel framework that combines learning of invariant descriptors through data augmentation and orthographic viewpoint projection. We propose rotation-robust local descriptors, learnt through training data augmentation based on rotation homographies, and a correspondence ensemble technique that combines vanilla feature correspondences with those obtained through rotation-robust features. Using a range of benchmark datasets as well as contributing a new bespoke dataset for this research domain, we evaluate the effectiveness of the proposed approach on key tasks including pose estimation and visual place recognition. Our system outperforms a range of baseline and state-of-the-art techniques, including enabling higher levels of place recognition precision across opposing place viewpoints, and achieves practically-useful performance levels even under extreme viewpoint changes. We reduce pose estimation error by 86.72% relative to state of the art. In particular, our integration of Visual Place Recognition (VPR) with SLAM by leveraging the robustness of deep-learnt features and our homography-based extreme viewpoint invariance significantly boosts the performance of VPR, feature correspondence and pose graph sub-modules of the SLAM pipeline. We demonstrate a localization system capable of state-of-the-art performance despite perceptual aliasing and extreme 180-degree-rotated viewpoint change in a range of real-world and simulated experiments. Our system is able to achieve early loop closures that prevent significant drifts in SLAM trajectories.

Abstract

With the number of protein sequences increasing rapidly, it becomes imperative to have a basic idea of the function and structure of a protein before the three-dimensional structure becomes available. Protein-drug interactions play essential roles in many biological processes and therapeutics. Prediction of the active binding site of a protein helps discover and optimise these interactions leading to the design of better ligand molecules. The secondary structure provides clues to the shape that the protein can be expected to take. It tells us whether an amino acid belongs to a coil turn, alpha-helix or beta-sheet structure. Deep Learning is a class of machine learning algorithms that progressively uses multiple layers to extract higher-level features from raw input. Deep learning methods eliminate feature engineering for supervised learning tasks by translating the raw inputs into intermediate representations that capture the more abstract and composite information, removing redundancies in the original input. The rapid adoption and success of deep learning algorithms in various sections of structural biology beckon deep learning algorithms for accurate binding site detection and secondary structure prediction. Protein-drug interactions play essential roles in many biological processes and therapeutics. Prediction of the active binding site of a protein helps discover and optimise these interactions leading to the design of better ligand molecules. The tertiary structure of a protein determines the binding sites available to the drug molecule. To quickly and accurately predict the binding site from sequence alone without utilising the three-dimensional structure is challenging. In the first study, a Residual Neural Network (leveraging skip connections) [1] is implemented to predict a protein’s most active binding site. An Annotated Database of Druggable Binding Sites from the Protein DataBank, sc-PDB [2], is used for training the network. Features extracted from the Multiple Sequence Alignments (MSAs) of the protein generated using DeepMSA, such as Position-Specific Scoring Matrix (PSSM), Secondary Structure (SS3), and Relative Solvent Accessibility (RSA), are provided as input to the network. A weighted binary cross-entropy loss function is used to counter the substantial imbalance in the two classes of binding and non-binding residues. The network performs very well on single-chain proteins, providing a pocket that has good interactions with a ligand. Secondary structure predictions predict three classes: C (Coil Turn), H (Alpha Helix) and E(Beta Sheet). In the second study, a Transformer (based on the multi-attention mechanism) network is used to train on the TR4590 dataset, which contains 4590 proteins with a sequence similarity cut off 25% and Xray resolution better than 2.0A. Ten models are trained across 10-fold cross-validation, and a weighted ˚ cross-entropy loss function is used. The ten trained models are run on the test set containing 1199 sequences. The mean of the probabilities of each class is taken, and then the class with the maximum probability is considered the class to which the amino acid belongs. The model achieves an accuracy of 82.51%.

Abstract

Transmission Control Protocol(TCP) is one of the most widely deployed transport layer protocol around the world. Since TCP is designed for reliable packet transfer, whenever the packet transfer is deemed unsuccessful, TCP takes various steps to resolve it such as re-transmitting the packet, changing the packet sending rate, etc. Congestion in the network is one such reason for unsuccessful packet transfer. The subsequent packet retransmission reduces the goodput in the network. There have been many variations of TCP for appropriate congestion control. These variations manage the sending rate of the users based upon the feedback they get from the network. In like manner, there are congestion control methods that leverage the facilities of the Software Defined Network(SDN) framework and manage congestion from the controller. The global image of the network in the controller enables the network administrators to deploy rules to manage the occurrences of congestion. In this work, we aim to couple both of these approaches. We make use of the reactive approach of the TCP congestion control at the user’s TCP stack. This helps us to manage congestion as it is detected at the user. To proactively reduce the possibility of the occurrence of congestion, we use a SDN controller to detect a congestion event even before it happens. We provide a utility based framework for this joint proactive and reactive mechanisms for congestion control. We propose a system optimization function based upon [15] and show that our user’s and network controller’s congestion control methods optimize the system utility. We simulate our model and show that our model significantly improves the system utility when compared with other congestion control methods

Abstract

Humans have used different characteristics to identify each-other since early times. This practice of identification based on person-specific features called biometric traits has developed over time to use more sophisticated techniques and characteristics like fingerprints, irises and gait in order to improve the identification performance. Fingerprints due to their distinctiveness, persistence over time and ease of capture, have become of the most widely use biometric traits for identification. However, with this ever increasing dependence on fingerprint biometrics, it is very important to ensure the safety of these recognition system against potential attackers. One of the most common and successful ways to circumvent these systems is through the use of fake or artificial fingers synthesized using commonly available materials like silicon and clay to match the real fingerprint of any particular person. Most fingerprint recognition systems employ a spoof detection module to filter out these fake fingerprints. While they seem to work well in general, it is a well-established fact that spoof detectors are not able to identify spoof fingerprints synthesized using ”unseen” or ”novel” spoof materials, i.e, the materials which were not available during the training phase of the detector. While it is possible to synthesize a few fingers using the various available materials, present-day spoof detectors require a large amount of samples for their training, which is practically not feasible due to the high cost and high complexity of fabrication of spoof fingers. In this thesis, we propose a method for creating artificial fingerprint images using only a very limited number of artificial fingers created from a specific material. We train a style-transfer network using available spoof fingerprint images which learns to extract material properties from the image, and then for each material, uses the limited set of spoof fingerprint images to generate a huge dataset of artificial fingerprint images without actually fabricating spoof fingers. These artificial fingerprint images can then be utilised by the spoof detector for training. Through our experiments, we show that the use of these artificially generated spoof images for training can improve the performance of existing spoof detectors over unseen spoof materials. Another major limitation of present-day recognition systems is their high resource requirements. Most fingerprint recognition systems use a spoof detector as a separate system either in series or in parallel with a fingerprint matcher leading to very high memory and time requirements during inference. To overcome this limitation, we explore the relationship between these two tasks in order to develop a common module capable of performing both spoof detection and matching. Our experiments show a high level of correlation between the features extracted for spoof detection and matching. We propose a new joint model which achieves similar fingerprint spoof detection and matching performance on various datasets as current state-of-the-art methods while using 50% less time and 40% less memory, thus providing a significant advantage for recognition systems deployed on resource-constrained devices like mobile phones.

Abstract

A wide variety of information pertaining to human behaviour, ranging from connections on social networks to accounts of discrimination and abuse, is being produced on the internet. Predictive models built using this data can be valuable for various domains including computational social science, personalization, recommendation, and marketing. Creating such predictive models can be challenging owing to the fact that the data available on online media relevant to human behaviour may be unstructured, inadequate and/or highly imbalanced. In this thesis, we consider two contrasting instances of human behaviour: trust and discrimination, especially gender-based discrimination, i.e., sexism. We formulate a novel approach based on matrix factorization for predicting user-user trust by capitalizing on community memberships and devising a new way to model the homophily theory. With regard to sexism, we introduce the problem of multi-label categorization of accounts describing sexism of any kind(s) and propose neural network based solutions for it. The prediction of trust relations between users of social networks is critical for finding credible information. Inferring trust is challenging, since user-specified trust relations are highly sparse and power-law distributed. In this thesis, we explore utilizing community memberships for trust prediction in a principled manner. We also propose a novel method to model homophily that complements existing work and a way of capitalizing on helpfulness scores of users’ item ratings. To the best of our knowledge, this is the first work that mathematically formulates a hypothesis pertaining to community memberships for unsupervised trust prediction. We propose and model the hypothesis that a user is more (or as) likely to develop a trust relation within the user’s community than (or as) outside it. Unlike existing work, our approach for encoding homophily directly links user-user similarities with the pair-wise trust model. Our formulation for tapping rating helpfulness scores relates to the notion that users receiving low rating helpfulness scores should not receive high levels of trust on average. We derive mathematical factors that model our hypothesis relating community memberships to trust relations, the homophily effect, and the premise involving helpfulness scores. Along with an existing method, they are combined into chTrust, the proposed multi-faceted optimization framework. Our experiments on the standard Ciao and Epinions datasets show that in the majority of the settings involving the fraction of the trust relations used for training, whether all user pairs or only the low-degree ones are considered in the evaluation, and the dataset, the proposed framework outperforms all but one of the unsupervised trust prediction baselines considered; no proposed or baseline method performs the best across all settings.We also develop a supervised method for trust prediction that avoids any dependence on auxiliary information such as ratings or reviews. We propose combining representation learning with the inference of trust relations. Using only a small amount of binary user-user trust relations, our approach simultaneously learns embeddings for the users and a trust prediction model. Another human behaviour that we endeavour to analyze through predictive models is discrimination, especially the form of discrimination that is perpetrated on the basis of sex. Sexism, an injustice that gives rise to enormous suffering, manifests in blatant as well as subtle ways. In the wake of growing documentation of experiences of sexism on the web, the automatic categorization of accounts of sexism has the potential to assist social scientists and policy makers in studying and thereby countering sexism. The existing work on sexism classification has certain limitations in terms of the categories of sexism used and/or whether they can co-occur. To the best of our knowledge, this is the first work on the multi-label classification of sexism of any kind(s), and we contribute the largest dataset for sexism categorization. We also consider the related task of the classification of misogyny. While sexism classification is performed on textual accounts describing sexism suffered or observed, misogyny classification is carried out on tweets perpetrating misogyny. We devise a novel neural framework for classifying sexism and misogyny that can combine text representations obtained using models such as BERT with distributional and linguistic word embeddings using a flexible architecture involving recurrent components and optional convolutional ones. Further, we leverage unlabeled accounts of sexism to infuse domain-specific elements into our framework. In order to evaluate the versatility of our neural approach for tasks pertaining to sexism and misogyny, we experiment with adapting it for misogyny identification. For categorizing sexism, we investigate m

Abstract

Computer Aided Diagnosis (CAD) systems are developed to aid doctors and clinicians in diagnosis after interpreting and examining a medical image. CAD systems aids in performing the task more consistently. With the arrival of data-driven deep learning paradigm and availability of large amount of data in the medical domain, CAD systems are being developed to diagnose a large variety of diseases ranging from different types of cancers, heart and brain diseases, Alzheimer’s disease, and diabetic retinopathy, etc. These systems are highly competent in performing the task on which they are trained. Although such systems perform at par with the trained clinicians, they suffer from a limitation in that they are completely black box in nature and are trained only on image-level class labels. This poses a problem in deploying CAD systems as stand alone solutions for disease diagnosis. This is because decisions in the medical domain are about health of a patient and are well reasoned and backed by evidence, sometimes from multiple modalities. Hence, there is a critical need for CAD system’s decisions to be explainable. Restricting our focus to only image modality, a solution to design explainable CAD systems could be to train the system using both class labels and local annotations and derive the explanation in a fully supervised manner. However, getting these local annotations is very expensive, time-consuming, and infeasible in most circumstances. In this thesis we address this explainability and data scarcity problem and propose two different approaches towards the development of weakly supervised explainable CAD systems. Firstly, we aim to explain the classification decision by providing heatmaps denoting important regions of interest in the image, which helped the model make the prediction. In order to generate anatomically accurate heatmaps, we provide a mixed set of annotations to our model - class labels for the entire training set of images and rough localization of suspect regions for a smaller subset of images in the training set. The proposed approach is illustrated on two different disease classification tasks based on disparate image modalities - Diabetic macular edema (DME) classification from OCT slices and Breast Cancer detection from mammographic images. Good classification results are shown on public datasets, supplemented by explanations in the form of suspect regions; these are derived using just a third of images with local annotations, emphasizing the potential for generalisability of the proposed solution.

Abstract

Quantum states that possess negative conditional von Neumann entropy (quantum conditional entropy) provide quantum advantage in several information-theoretic protocols including superdense coding, state merging, distributed private randomness distillation and one-way entanglement distillation. While entanglement is an important resource, only a subset of entangled states have negative conditional von Neumann entropy. This thesis deals with the characterization, study and detection of these resourceful states. The first part of the thesis deals with the detection of negative conditional entropy states. We characterize the class of density matrices having non-negative conditional von Neumann entropy as convex and compact. This allows us to prove the existence of a Hermitian operator (a witness) for the detection of states having negative conditional entropy for bipartite systems in arbitrary dimensions. We show two constructions of such witnesses. For one of the constructions, the expectation value of the witness in a state is an upper bound to the conditional entropy of the state. We pose the problem of obtaining a tight upper bound to the set of conditional entropies of states in which an operator gives the same expectation value. We solve this convex optimization problem numerically for a two qubit case and find that this enhances the usefulness of our witnesses. We also find that for a particular witness, the estimated tight upper bound matches the value of conditional entropy for Werner states. We explicate the utility of the methods proposed in the detection of useful states in several protocols. In the second part of the thesis, we introduce the resource-theory framework, and give pertinent directions for the study of conditional entropy under the resource-theoretic framework.

Abstract

Water is an essential resource for sustaining human life. The ever-increasing demand for clean water from the growing human population and other anthropogenic needs have pushed several regions into highly water- stressed areas. Diminishing freshwater resources and high amounts of water contamination are increasing the threat of severe water scarcity even further. Since the total available freshwater resources cannot be modified, the options are to maintain and/or improve water quality in water bodies and increase the reusability of the water. The former can be achieved by continuous monitoring of the water quality in the water bodies and identifying and regulating the source of contaminations that pollute these water bodies. So, water quality in rivers and large inland water bodies are being measured occasionally through in-situ approaches. While these do provide valuable inputs, they are not sufficient to understand the underlying causes for such changes in water quality, including the seasonal and other drivers of contamination. Hence, there is a need to monitor water bodies, especially the lakes and reservoirs, regularly. Constant inflows and long storage time of water with limited or no proper outflow make these water bodies much more susceptible to the adverse effects of contaminants than rivers. This study takes a three-pronged approach to study and understand the water quality in inland water bodies – first, by developing and evaluating a methodology for monitoring nutrient contamination and its changes in inland water bodies using remote sensing satellite data; second, to hypothesize and study the probable source of the contamination as land use and its practices; and three, to understand, analyze and model the interactions between the land use changes in the contributing watershed and the water quality to help improve the regional- scale decision making capabilities. The case study area is Nagarjuna Sagar (NS) reservoir in the Krishna River basin, one of the largest inland water bodies in India. NS is a multipurpose dam and inland water body (reservoir) with a spatial spread of 285 Km2 and a catchment area of 215,000 Km 2 . In addition to Irrigation and Power generation, it is also a primary source of drinking water to Hyderabad, a large metropolis in India with nearly 10million population. All these make it essential to study and maintain the water quality of the NS water body. While the literature suggests that the remote sensing approach can help monitor the nutrient content in the water body, most studies and analyses have been done for oceans, seas, and enormous water bodies. Their application to inland water bodies is still a challenge due to the spatial and spectral limitations of the remote sensing sensors. Hence, the study explored the effectiveness of the remote sensing data in studying the inland water body nutrient contamination. Here, the presence and spatial spread of Chl-a, a remote sensing detectable agent, present in the photosynthetic organisms such as algae, phytoplankton, and cyanobacteria (AP) was used as a proxy for monitoring nutrient contamination in water bodies. The unique spectral signature of Chl-a was used to detect and identify its spatial spread across the water body. For this process, a decision-tree-based classification technique was developed that exploits the spectral response peaks within the region of 705nm to 860nm for each pixel to detect and classify them into No to Low contamination, Moderate and High Chl-a levels. The spatial spread of moderate and high Chl-a areas was used to understand the severity of the contamination in the water body. Further, through meta-analysis and comparative study, it was found that MODIS and Sentinel satellite data are best suited for the detection of Chl-a content in the water body using the spectral signature-based method. While various works have well-studied MODIS data, Sentinel data is relatively recent and needs to be tested for its sensitivity towards Chl-a across various concentrations of Chl-a in water bodies situated in different environments. Since the NS water quality data is not available, it is prudent to use the water quality data from earlier published studies to evaluate the method. As part of this, Lake Taihu (TL, with an area of 2250 Km 2 ) and Lake Bebe (LB, with an area of 24 Km 2 ) with different concentrations of Chl-a were used to test the sensitivity of the method. The Manasarovar Lake (ML) (area of 411 Km 2 ) was used as a zero baseline case as it is free of any Chl-a contamination to check if the method was working without any false positives. The results from TL showed that more than 50% of the water body indicates the presence of Chl-a content which matched with the data given in the literature. Similarly, the method could not detect any Chl-a content in the ML, which agrees with the literature but showed some limitations in the case of LB. On applying

Abstract

With rapid advancements in technology, the volume of data generated keeps increasing at an incredible pace. Data mining is the process of discovering novel and potentially useful information from large amounts of data. Currently, data mining techniques are being employed by organizations to improve performance by extracting useful knowledge from large datasets. Pattern mining is one of the important data mining tasks. The pattern mining techniques are being applied in applications such as social media, e-commerce, advertising and so on. In the literature, models and algorithms have been proposed for mining different kinds of patterns such as frequent patterns, correlated patterns, utility patterns, weighted frequent patterns and coverage patterns. There is an ongoing research to develop efficient parallel algorithms to extract patterns from large datasets by employing parallel processing frameworks such as shared memory, shared nothing, GPUs, Grid computing and MapReduce. In this thesis, we proposed parallel algorithms by employing MapReduce framework to extract correlated patterns and weighted frequent patterns. Notably, MapReduce framework is a programming model that enables distributed processing of huge amounts of data on a very large number of commodity machines in geographically distributed environments in a reliable and fault-tolerant manner. Once an algorithm is developed on a MapReduce framework, it can be easily scaled to process large datasets by employing a large number of machines in a shared-nothing environment. The correlated pattern mining involves the extraction of patterns from transactional databases subject to the user specified minimum support and minimum all-confidence constraints. Typically, for any parallel algorithm that runs on multiple machines, the data is segmented into multiple partitions and distributed among the machines. This segmentation influences the data shuffle among the machines and runtime of the algorithm. We introduce two novel data segmentation techniques - database segmentation and transaction segmentation to discover the correlated patterns efficiently. The former technique involves segmenting the database into multiple sub-databases such that each sub-database can be mined independently. The latter technique involves segmenting a transaction into multiple sub-transactions such that each sub-transaction can be processed as an individual transaction. We employ these data segmentation techniques and leverage MapReduce framework to propose Parallel Correlated Patterngrowth algorithm.The proposed segmentation techniques are algorithm agnostic, and therefore, can be incorporated into any parallel correlated pattern mining algorithm. The process of weighted frequent itemset (WFI) mining involves the extraction of itemsets from a given transactional database subject to the user defined minimum support, minimum weight and min imum weighted support constraints. It has the following two challenges: (i) finding WFIs is a computationally expensive process as these itemsets do not satisfy the downward closure property and (ii) there are only few computationally efficient algorithms to find WFIs in very large databases. Firstly, we propose Sequential Weighted Frequent Pattern-growth algorithm to discover WFIs efficiently. It employs three novel pruning techniques to reduce the computational cost effectively. We also leverage the MapReduce framework to achieve further computational optimization and propose Parallel Weighted Frequent Pattern-growth algorithm. Overall, we propose MapReduce based approaches to extract correlated patterns and weighted frequent patterns from large transactional databases. We demonstrate that the proposed approaches are highly scalable, memory and runtime efficient through results of experiments conducted on both realworld and synthetic datasets.

Abstract

Speech produced in practical environments is affected by several sources of degradations, which includenoises, reflections, reverberation and sounds from other sources such as background speech, vehicles, etc. Thesedegradations reduce the performance of several speech-based applications, such as speech and speaker recogni-tion systems. It is difficult to characterize these degradations, as they are not known in advance, and also theyare time varying. Hence, it is necessary to explore and extract speech-specific and speaker-specific informationin the collected speech signal to improve the performance of speech-based systems. The objective of the studiesproposed in this thesis is to explore methods based on single frequency filtering (SFF) analysis of speech signalsand artificial neural network (ANN) models for some speech-specific and speaker-specific tasks. The SFF analy-sis gives flexibility in the representation of speech information at different spectral and temporal resolutions. TheANN models are used for capturing the implicit features, specific to a given application. The SFF analysis can beused to provide a representation of the excitation source and the vocal tract system characteristics of the dynamicspeech production mechanism. The ANN models help in extracting the discriminative and descriptive featuresas needed for a given task. The speech-specific tasks explored in this thesis include speech/nonspeech detectionfrom degraded speech, discrimination of natural and synthetic speech. The speaker-specific tasks examined inthis thesis include speaker separation in two speaker conversation data, mimicked voice detection and speakerverification.The speech/nonspeech detection task is addressed using approaches based on signal processing and on ANNmodels. The signal processing approach exploits speech-specific characteristics in the degraded signal. The re-sults of speech/nonspeech detection are given for the speech data from TIMIT corpus, corrupted by syntheticallyadding noises of different types at different levels. The signal processing approach is also used for detectingthe speech regions, when the speech is corrupted by unknown sources of degradations in a practical informationaccess situation. The performance of speech/nonspeech discrimination using ANN models is examined for syn-thetically added degradations to speech signals, mainly to study the effectiveness of different types of SFF-basedrepresentations of the signal. These models have been successfully used for voice activity detection (VAD) taskon the naturalistic Apollo corpus. The effectiveness of SFF-based representations and the discriminative charac-teristics of ANN models are examined for several speech-specific tasks such as discrimination of close vs distantspeech, live vs recorded speech, and natural vs synthetic speech.Speaker-specific tasks such as study of speaker separation, mimicked voice detection and speaker verificationare explored using discriminative and descriptive models of ANN. Speaker separation task is successfully demon-strated on several cases of 2-speaker conversation data. The proposed approach for speaker verification task isdemonstrated only for a limited number of speakers, as the speaker verification task should not be dependenton the number of speakers. Overall, this study examined the effectiveness of different SFF-based representa-tions for some speech-specific and speaker-specific tasks using ANN models. It also highlighted deficiencies inrepresentation for some other tasks.