Abstract

Glaucoma is a major eye disease which when untreated, can gradually lead to irreversible loss in vision. The underlying causes are a loss of retinal nerve fibres (resulting in a thinning of the layer and enlargement of the optic cup) and peripapillary atrophy. Since these occur without any sign of symptoms in the initial stages, it is difficult to diagnose the disease in the early stages. Hence, the development of Computer-aided diagnostics (CAD) systems for early detection and treatment of the disease has attracted the attention of many medical experts and researchers alike. Optical Coherence Tomography (OCT) and Fundus photography are two widely used retinal imaging techniques for obtaining the structural information of the eye which helps to analyze and detect the diseases. Existing automated systems rely largely on fundus images for assessment of glaucoma due to their fast acquisition and cost. OCT images provide vital and unambiguous information for understanding the changes occurring in the retina, specifically related to the retinal nerve fiber layer and the optic nerve head which are essential for disease assessment. However, the high cost of OCT is a deterrent for deployment in screening at large scale. Hence, the focus of this thesis is to investigate the potential of integrating the two retinal imaging techniques which provide complementary information of the eye for developing automatic glaucoma screening system. Firstly, we propose a deep learning approach directly operating on 3D OCT volumes for glaucoma assessment which showed promising results, thus demonstrating the prominence of the highly discriminative features learnt from OCT for automated glaucoma detection. Next, we present a novel CAD solution wherein both OCT and fundus modality images are leveraged to learn a model that can perform a mapping of fundus to the OCT feature space. We show how this model can be subsequently used to detect glaucoma given an image from only one modality (fundus), thus enabling the automated screening operation to be executed on a large scale. The results show that fundus to OCT feature space mapping is an attractive option for glaucoma detection

Abstract

Language identification (LID) refers to the task of identifying the language being spoken by a speaker in a given speech signal. LID system acts as a front-end module for different applications such as multilingual automatic speech recognizer, multilingual dialogue system, and voice service systems. Notably, in an Indian scenario, where almost every state has a language of its own and every language having various dialects, development of a LID system becomes crucial. In literature, the standard acoustic features such as Mel-frequency cepstral coefficients, shifted delta cepstral coefficients and prosody (pitch, duration, and intonation) have been explored in the Indian context. These features cannot adequately discriminate the languages since most of the Indian languages have an overlapped set of phonemes. Even though Indian languages have an overlapped set of phonemes, due to phonotactic constraints, the characteristics of a particular sound unit differ in different languages which motivates us to explore phonetic features. In the literature, most of the works used senone based deep neural networks to extract phonetic features. We have used a joint acoustic model (JAM) which is trained using long short term memoryconnectionist temporal classification (LSTM-CTC) network to extract phonetic features. The JAM implicitly learns the language-specific information without any prior knowledge of the language. Hence, it is hypothesized that features extracted from JAM effectively discriminate the languages. These features are referred to as CTC features. The temporal variations in CTC features are modeled using the attention-based residual time-delay neural network to predict language ID. Attention mechanism aggregates frame-level features by selecting prominent frames through a parametrized attention layer. The performance of the proposed LID framework (CTC with attention-based RES-TDNN) has been evaluated on IIITH-ILSC database, which consists of 22 official Indian languages and Indian English. The proposed network outperformed other state-of-the-art methods with an equal error rate of 8.82%.

Abstract

“Natural language processing is the “big data” problem for the future of computing. While a huge amount of work has been done in this area, which we call ‘modeling and parameterizing computer systems,’ a fundamentally different approach is needed. We want to deploy virtual languages, and human computational agents. We have no formal computational theory that describes the problems that will be addressed by these. Most existing work is preliminary, although some work has been done and some actual software exists.” Can you guess the author who penned the above quote? You’ll be surprised it is not a famous celebrity. No, it is not even an overworked computer scientist. In fact, the entire quote was generated automatically by OpenAI’s latest language model, GPT-2 on simply supplying a short prompt of “Natural Language Processing is”. Surprising, isn’t it? Indeed, that’s how powerful language models can be! Language models are at the heart of Natural Language Processing (NLP). They help in extracting regularities from natural language which aid in solving multiple problems in various applications like automatic speech recognition, document classification, sentiment analysis, summarization, et cetera. With the growing advent of deep learning techniques aided by huge and efficient computational resources - we seek to evaluate various applications of fine-tuning language models which can guide us to a better and nuanced understanding of natural language. In this thesis, I’ll attempt to tackle some of the original problems that language models can help solve. Primary among those is the style transfer task wherein a piece of text is desired to be stylistically rewritten in the fashion of a given target author. Due to the unavailability of parallel data for such a task, it is not possible to leverage supervised systems like sequence-to-sequence models which entirely rely on paired instances of training data. This is where we bring in the concept of language-model finetuning - wherein we build a novel stylistic language model, StyleLM. StyleLM can be used to rewrite any given piece of text into the style of a target author, and is also scalable to multiple authors. We demonstrate the ability of StyleLM to rewrite into the style of a given target author across 3 different domains - viz. online reviews, fiction books and encyclopaedic pages. Our evaluation framework shows that StyleLM achieves high alignment with the target authors across all 3 levels (surface, lexical, syntactic) of the stylistic spectrum. At the same time, StyleLM also achieves comparable performance to the supervised baseline in both qualitative and quantitative comparisons. We also seek to evaluate the power of language models across content understanding tasks - both at the syntactic and semantic levels of linguistic discourse. For syntactic understanding, we dive into search queries, and seek to evaluate if pretraining and finetuning large language models can help discern wellformed search queries effectively from the non well-formed ones. To effectively understand the semantic spectrum, we use data from happy moments posted by users online to determine if we can detect the agency (whether the author is in control?) and social (if anyone other than the author is involved?) characteristics from within the happy moments. Across both these tasks, we achieve considerably high performance - with the latter system for understanding happy moments achieving the runner-up at the shared task in an affective content analysis workshop held at AAAI 19. We also outperform the stateof-the-art for syntactic discernment of search queries.

Abstract

Understanding the myriad interactions between protein and DNA that are present at the heart of most important biological phenomena is of critical importance to understanding life itself. Molecular Dynamics (MD) simulations provide a detailed view of the conformational changes that drive these intricate, multi-stage, highly dynamical protein-DNA binding processes. Aided by MD, the present work aims to investigate two processes that DNA-binding proteins take part in: (a) recognizing DNA mismatch lesions to commence the repair mechanism, and (b) regulating and affecting the process of transcription. The ultraviolet (UV) radiation-induced DNA lesions play a causal role in many prevalent genetic skin-related diseases and cancers. The damage sensing protein Rad4/XPC specifically recognizes and repairs these lesions with high fidelity and safeguards genome integrity. Despite considerable progress, the mechanistic details of the mode of action of Rad4/XPC in damage recognition remain obscure. The present thesis investigates the mechanism, energetics, dynamics, and the molecular basis for the sequence specificity of mismatch recognition by Rad4/XPC. We dissect one of the key molecular events that occur as Rad4/XPC tries to recognize and bind to DNA lesions/mismatches: the flipping of a pair of nucleotide bases at the damage/mismatch site. Using suitable reaction coordinates, the free energy surfaces for the base flipping event were determined using molecular dynamics (MD) and umbrella sampling simulations on three mismatched (CCC/CCC, TTT/TTT and TAT/TAT mismatches) Rad4-DNA complexes. We identify the key determinants of the sequence-dependent specificity of Rad4 for the mismatches and explores the ramifications of specificity in the process of base flipping. The results unravel the molecular basis for the high specificity of Rad4 towards CCC/CCC mismatch and lower specificity for the TAT/TAT mismatch. The interplay of the conformational flexibility of mismatched bases and the Rad4 specificity explored here complement recent experimental FRET studies on Rad4-DNA complexes. The cAMP-mediated allosteric transition in the catabolite activator protein (CAP) allows it to position its recognition F-helices to bind to the promoter region on DNA, resulting in gene activation. The present study aims to provide a molecular basis for the conformational changes observed during the allosteric regulation of CAP by cAMP and the subsequent binding to the DNA promoter region using MD simulation and umbrella sampling techniques. In particular, we investigate DNA binding-induced changes in the orientations and dynamics of functionally-relevant F-helices in the DNA binding domain. The two-dimensional free energy profiles of the cAMP-liganded CAP and the cAMP-CAP-DNA complex provide a clear, simple picture of the flexibility constraints imposed on the DNA binding region of CAP upon DNA-binding. We further relate these differences in the free energy surface of CAP to the dynamical modes of the recognition F-helices. The observed dynamical changes are then explained by dissecting the key interactions between CAP, DNA, and the cAMP molecules.

Abstract

We are given an environment with some objects (park with trees, a city block area) and mobile agents moving in the environment. The agents have a limited field of view and visibility range. An agent (hider) can hide behind an object to be not seen by other agents (seekers) through their line of sight (visibility). These hider agents (that are hiding behind objects) can be found by seekers who explore the environment. A pragmatic example of the problem is a warehouse environment or a cordoned off area of the town, with no drones or cameras, and some culprit agents are hiding from the police agents. The aim of hiders is not to be caught for the longest time, and the aim of the seekers is catch all of them in the shortest period of time. We formulate the problem by using visibility based map abstractions. These abstractions involve the notions of coverage, obstruction and escapability. Using these spatial abstractions as basic building blocks, we formulate two kinds of strategies for the agents. The first kind is based on multi-armed bandits in which the agents maintain reward bounds on obstacle edges which they feel are safe to hide or suitable for searching. The second kind focuses on deterministic capture of hiders, even if they are guided by an oracle which knows the future positions of seekers. We built a simulator to create environments for multiple hiders and multiple seekers to test their strategies in a Hide-and-Seek game. We contrast our strategies, and their variants, by empirically comparing their game completion times.

Abstract

Modern times have witnessed unprecedented material growth leading to increase in efficiency, comfortability and longevity of life. The success of modernity has attracted attention and curiosity of global community. However, behind the silver screen of modern glamour lie some existential challenges which are threatening not only the human welfare but also the existence of life itself on this planet. Some of these challenges include ever growing unsustainability of modern lifestyles, decimation of diversity in various spheres and a general disinterest seen in modern mind towards issues which do not have immediate bearing on one’s living (specifically material wellbeing). These challenges ought to be taken up head on in contemporary times in order to address the concerns fundamentally. In this context, exploration of the jajmāni system which was prevalent in traditional Indian society appears to be an important domain for study. Indian society has been known for being home to rich diversity. India has been home to all the major religions of the World, hundreds of languages are spoken by its people, thousands of gods are worshipped, and many important philosophical schools find patronage in Indian social institutions. Also, there are sufficient indications that there used to be a prime concern as well as facilitation of ‘collective contemplation’ on fundamental issues of humanity including the primal condition of man through various artistic modes. The jajmāni system also deserves to be analyzed from the point of view of sustainability and local self-sufficiency as it is usually claimed that it provided widespread prosperity to people. Even though the jajmāni system has attracted the interest of large number of scholars in the last century, however not much emphasis is discernible in these works has been done to understand these concerns. The present study on jajmāni system has attempted to study it from the following four major concerns The birth and sustenance of diversity in jajmāni villages - The role and importance of ‘culture of collective contemplation’ with regard to ensuring ‘unity-in-diversity’ in the society. - The self-sufficiency and sustainability of jajmāni villages. - And, the causes behind the collapse of jajmāni system in modern times. For all these, it is also important to grasp the philosophical basis and the key principles on which the jajmāni system was established over a long period. However, it is pertinent to note that at present it is not possible for one to come across a village which is actually functioning on the principles of jajmāni system. Therefore, for a comprehensive understanding, a conceptual model has to be visualized based on its formative principles and the bits of available information on it. In this regard, the present study reviewed available literature to identify and understand the underlying philosophical concepts and also to understand the system’s functioning. We have also attempted to identify and learn from a few rather unconventional resources to improve upon our understanding. One important resource has been Shri Ravindra Sharma, who extensively studied the functioning of the system in Adilabad region of southern India by actually living amongst several traditional communities for more than two decades and learning their crafts. In addition to him, this study has also extensively interacted with community elders of two communities in Western India- the kurbas (shepherds) and the thākar (traditional puppeteers). The spectrum of our literature review has addressed two broad interests- one, to understand salient concepts that are needed for this study and the other, to critically assess the existing status of information on jajmāni system. Consequently, the whole review is categorized in four different but connecting sections. The four seThe insights obtained during the course of this study can be useful to address the problem of diversity sustenance in contemporary society and the challenges of sustainability and self- sufficiency for local community living.ctions are as follows: - Salient views on Origins of ‘Society’ and Concerns of Diversity - The Importance of Collective Contemplation and the roles of Public-Private Realms - A critical review of the core literature on jajmāni system - Studies to understand philosophical basis of jajmāni system The jajmāni system in very brief can be understood as a social system of direct inter-dependency and mutual fulfilment amongst the families in a village community. The institution of jāti (caste) becomes pivotal to this arrangement. The system is founded on the basic philosophical principlesof mokśa (liberation)- as the purpose of life and ‘yajña’ (mutual fulfillment) as a way of living. The diversity in the jajmāni system is protected and sustained by the institution of jāti (caste), which ensures family-based production system. Moreover, diversity gets constant

Abstract

Time-sharing, which allows for multiple users to use a shared resource, is an important and fundamental aspect of modern computing systems. However, accelerators such as GPUs, that come without a native operating system do not support time-sharing. The inability of accelerators (such as GPUs) to support time-sharing limits their applicability especially as they get deployed in Platform-as-a-Service (PaaS) and Resource-as-a-Service (RaaS) environments. In the former (PaaS), elastic demands may require preemption where as in the latter (RaaS), fine-grained economic models of service cost can be supported with time-sharing. We extend the concept of time-sharing to the GPGPU computational space using cooperative multitasking approach. Our technique is applicable to any GPGPU program written in Compute Unified Device Architecture (CUDA) API provided for C/C++ programming languages. With minimal support from the programmer, our framework incorporates process scheduling, light-weight memory management, and multi-GPU support. Our framework provides an abstraction where, every workload can use a GPU(s) over a time quantum exclusively. We demonstrate the applicability of our scheduling framework, by running many workloads concurrently in a time-sharing manner. In this context, we explore two scheduling alorithms: Round-Robin(RR) and Shortest Job First(SJF). We demostrate the versatile nature of the framework by running workloads from different domains.

Abstract

Scene Understanding has been a major aspiration of computer vision from its early days. Its root lies in enabling the computer/robot/machine to understand, interpret and manipulate visual data, in similarity to what an average human eye does in front of a natural/artificial localized location/scene. This ennoblement of the machine have a widespread impact ranging from Surveillance, Aerial Imaging, Autonomous Navigation, Smart Cities and thus scene understanding have remained as an active area of research in the last decade. In the last decade, the scope of problems in the scene understanding community has broadened from Image Annotation, Image Captioning, Image Segmentation to Object Detection, Dense Image Captioning, Instance Segmentation etc. Advanced problems like Autonomous Navigation, Panoptic Segmentation, Video Summarization, Multi-Person Tracking in Crowded Scenes have also surfaced in this arena and are being vigorously attempted. Deep Learning has played a major role in this advancement/development. The performance metrics in some of these tasks have more than tripled in the last decade itself but these tasks remain far from solved. Success originating from deep learning can be attributed to the learned features. In simple words, features learned from a Convolutional Neural Network trained for annotation are in general far more suited for captioning then a non-deep learning method trained for captioning. Taking cue from this particular deep learning trend, we dived into the domain of scene understanding with the focus on utilization of prelearned-features from other similar domains. We focus on two tasks in particular: Automatic (multi-label)Image Annotation and (Road)Intersection Recognition. Automatic image annotation is one of the earliest problems in scene understanding and refers to the task of assigning (multiple) labels to an image based on its content. Whereas intersection recognition is the outcome of the new era of problems in scene understanding and it refers to the task of identifying an intersection from varied viewpoints in varied weather and lighting conditions. We focused on this significantly varied task approach to broaden the scope and generalizing capability of the results we compute. Both image annotation and intersection recognition pose some common challenges such as occlusion, perspective variations, distortions etc. While focusing on the image annotation task we further narrowed our domain by focusing on graph based methods. We again chose two different paradigms: a multiple kernel learning based non-deep learning approach and a deep-learning based approach, with a focus on bringing out contrast again. Through quantitative and qualitative results we show slightly boosted performance from the above mentioned paradigms. The intersection recognition task is relatively new in the field. Most of the work in field focuses on Places Recognition which utilized only single images. We focus on temporal information i.e. the traversal of the intersection/places as seen from a camera mounted on a vehicle. Through experiments we show a performance boost in intersection recognition from the inclusion of temporal information

Abstract

Speech signal can be considered as a sequence of acoustic events, where each acoustic event can be distinctly characterized by its production mechanism. Acoustic event-based signal pro- cessing techniques have gained vital importance in speech feature representation which is ro-bust to noise. The detection of specific acoustic events that correspond to the unique phonetic features of the speech is important in speech assessment systems and clinical speech analysis.In landmark or acoustic event based speech analysis, the speech signal is processed around specific locations known as events instead of the entire speech signal. Most of the Indian languages are vowel-centric, and hence consonants must be grouped with a vowel following it or preceding it. This process of clustering consonant or consonant cluster with either preceding vowel or next vowel is known as syllabification. The significance of accurate vowel region detection from a continuous speech has many applications that include robust speaker recognition, smart audio filtering, consonant-vowel unit recognition in Indian languages, speech rate manipulation and multimedia synchronization. With this motivation, the thesis explores the use of speech production knowledge in the detection of vowel regions. Further, the vowel region detection techniques are used to detect retroflex approximants and palatalized consonants in the Indian context. The important issues addressed in this thesis are listed below:• A two-stage algorithm is proposed to detect precise vowel regions using dominant spectral peaks of zero frequency filtered speech signal. In post-processing, spurious vowel regions removed based on uniformity of epoch intervals and the positions of vowel on set points, and vowel end-points are corrected using the strength of the excitation of the speech signal.• A technique is proposed for the vowel region detection from the continuous speech using an envelope of the derivative of the speech signal, which is a non-negative, frequency weighted energy operator.• A language identification system using deep neural networks with attention is explored with a vowel region detection scheme incorporated at the front-end. The spectral features extracted from the vowel regions instead of entire speech are used to model the language identification system. • An approach for automatic detection of retroflex approximant in a Tamil continuous speech is proposed using vowel regions and formant dynamics. The formant dynamics are obtained through the Hilbert envelope of the numerator of the group delay function. • A technique is proposed to detect auxiliary palatalization from the continuous Kashmiri speech. These consonants investigated in synchronous with vowel regions, which are spotted using the instantaneous energy computed from the envelope-derivative of the speech signal. • A new speech feature representation is proposed by integrating single frequency filtering technique with higher order nonlinear energy operator for emotion recognition. The performance of the system is evaluated by extracting the speech features in vowel regions instead of entire speech in noisy environments.The final remarks drawn out of this thesis work are as follows: The performance of the proposed vowel region detection methods observed to be better compared to the existing meth- ods under clean and noisy environments. The proposed techniques utilize speech production knowledge in terms of strength of the excitation and uniformity of the epoch intervals of the speech signal. Moreover, vowel region detection technique based on the envelope of the derivative of the speech signal is better compared to the method based on the zero frequency filtering technique. It is mainly due to the capturing of the instantaneous energy of the speech signal in a better manner. A front-end vowel region detection scheme incorporated to the language identification system to use spectral features extracted from the vowel regions instead of the entire speech signal. It exhibited a significant improvement in the recognition rate in noisy environments. A speaker independent retroflex approximant detection approach is proposed to detect the language-specific phonetic feature of Tamil in clean and noisy environments. In an- other study, palatalization in Kashmiri is detected by anchoring vowel regions. The technique for detecting palatalization process in the vowel context is proposed based on set of acous- tic cues. These include a second formant gradient, a measure of the second formant nearing the third formant, and high to low energy ratio derived from the spectral characteristics of the palatalized consonants using the Hilbert envelope of the numerator of the group delay function. The results showed that these acoustic cues capture secondary articulatory changes in palatalized consonants to detect them in continuous speech. Lastly, a new feature representation is proposed for emotion

Abstract

Natural Language Processing deals with the techniques and processes which help computers interpret and process the human languages. There are various NLP applications which help us in daily tasks,namely spellchecker, text auto-correction, machine translation systems, text classification, and text summarization etc.. There is vast information available on the internet, but most of it is available in English.When it comes to Indian languages, the content is much lesser compared to other languages. Most of the Indian people are multilingual and machine translation systems can be utilized to translate from text in one language to a one’s own language. In that way machine translation technology can be helpful to overcome the language barrier. Also, most people in India are not proficient in English, hence machine translation systems can take advantage of the large information available on the internet. The machine translation technology has various approaches out of which main approaches are rule based, data driven and hybrid approach. Each of these approaches have their pros and cons. The hybrid approach utilizes the benefits of both rule based as well as data driven approaches by combining them. The pipeline based hybrid machine translation systems exhibit certain software engineering charac- teristic like heterogeneous nature of architecture, compute and knowledge intensive nature of modules and blackboard architecture. Due to these characteristics, the development and re-engineering of the hybrid MT systems are different from that of traditional software engineering processes. The hybrid MT systems with pipeline based architecture face some issues like lack of efficiency, architecture mismatch, scalability, robustness and performance related challenges in their integration, deployment and engineering. We took Anusaaraka (English to Hindi) and Sampark (Indian language to Indian language) MT systems as our use-cases of pipeline based hybrid machine translation systems. The different modules in both of the systems are developed by different teams. By studying both of the systems, we observed that modules, internally, sometimes use heterogeneous utilities and platforms as well as hard coded knowledge sources which makes them rigid in nature and difficult to manage. This is not in consonance with the larger specifications of the system. These modules are comprised of complex, knowledge and compute intensive NLP modules relying on a bulk of hard-coded knowledge sources. Such modules are typically very unwieldy when we expect to scale the performance of the overall NLP system they are a part of. Additionally, there are NLP applications which either utilize MT systems or support their development processes. Such applications are translation workbenches and resource management systems. We also studied the aspects of MT related applications and found that the translation workbenches as well as resource management systems have performance issues and are developed with heterogeneous modules as mentioned above. Therefore, for making MT systems as well as MT related NLP applications to be utilized by the mass public, these MT applications need to be re-engineered. In this thesis, we aim to propose an approach to re-engineer MT and MT related applications to make them efficient and improve their performance in terms of processing speed. It would facilitate the MT systems to be used as services which can be used in various NLP products and services. We first studied the Anusaaraka MT system and identified the NLP components which can be re-engineered. We proposed the symbiotic software engineering approach followed by micro-services based architecture to re-engineer the system. In this way, each of the NLP modules as well as the complete MT system is proposed as a set of independent microservices. We demonstrate our approach on Anusaaraka MT system in which the performance of the system improves by 87% on execution of 1 sentence. The documentation and test cases are also being prepared for each of the modules as well as the complete system. We also propose the concept of distributed caching which is utilized to improve the performance of compute intensive NLP modules in distributed environment. We performed experiments to analyze the impact of distributed caching on the Sampark MT system. The performance of the Sampark MT system improves additionally by 4-6% when we utilize distributed caching mechanism over normal caching. We then propose the translation workbench-Transzaar with a resource management system- Kunji to aid the translators. The performance of the translators is evaluated in various scenarios when they use Transzaar and Transzaar with resource management system-Kunji. The productivity of the translators improved significantly by 1.35-1.45 times for English-Hindi language pair and 2 times for Hindi-Urdu language pair to that of their initial productivity with Transzaar. Moreover it is also observed th