Abstract

India’s rich linguistic diversity, encompassing 22 scheduled languages and thousands of regional varieties, poses significant communication challenges across domains such as education, healthcare, governance, and digital access. Addressing this linguistic divide is crucial for fostering social and economic inclusion. Speech-to-Speech Machine Translation (SSMT) offers a promising solution, enabling seamless multilingual communication. However, developing SSMT for Indian languages is challenging due to the scarcity of corpora and the complexity of linguistic structures. This work adopts a cascaded SSMT approach, integrating automatic speech recognition (ASR), machine translation (MT), and text-to-speech (TTS), with a focus on overcoming the challenges of error propaga- tion between components to enhance translation quality. The primary focus is on enhancing the machine translation component while addressing critical gaps between automatic speech recognition and MT. These gaps include missing punctuation, spoken disfluencies, and grammatical inconsistencies, etc. We also ensure effective integration with text-to-speech systems by automatically generating subtitle text that aligns with video/audio content based on the source language. To achieve these goals, we proposed and developed a speech-to-speech or video-to-video translation pipeline incorporating several pre-processing and post-processing tools to bridge the above discussed gaps. These tools include punctuation restoration, disfluency detection, domain and domain-term identification, automatic translated subtitle generation, and video syncing, etc., supporting both English and Indian lan- guages. To improve MT quality across English and multiple Indian languages, we explored linguistic features such as part-of-speech (POS) tags, chunk tags, and morphology, with a particular emphasis on low-resource languages. We addressed the scarcity of parallel corpora for 36×36 language directions, including English and 35 Indian languages, by employing advanced data augmentation techniques such as iterative back-translation, discourse back-translation, COMET-based filtering, and pivot translation, resulting in the creation of over 10 billion parallel sentences. Additionally, we developed domain-specific human-translated parallel corpora tailored for general, educational, and healthcare domains. We also addressed translation evaluation by creating millions of data for direct machine translation evaluation and used them to develop both reference-free and reference-based evaluation metrics, leveraging encoder-decoder and decoder-only models. We also explored error identification and automatic post-editing as tasks related to MT. Our work further encompasses subtitle translation and video-to-video synchronization, ensuring precise alignment between translated subtitles and video frames.A significant contribution of this work is the development of a single multitask model with 2.5 billion parameters, integrating translation, discourse machine translation, evaluation, error detection, and post-editing functionalities. Additionally, we designed a decoder-only large language model tailored for English and 35 low-resource languages as a foundational model, which we further fine-tuned for MT and related tasks. We developed benchmark datasets for Speech-to-Speech Machine Translation (SSMT), disfluency translation, discourse machine translation, and domain-specific machine translation. Through these efforts, we present a scalable and robust SSMT system that fosters linguistic inclusivity and bridges communication gaps across India’s diverse languages. All our contributions, including corpora, models, and related resources, are made openly available to the community to facilitate further research and innovation.

Abstract

Urban air quality monitoring faces a fundamental scalability crisis. With only 800 monitoring stations serving over 4,000 Indian cities, traditional sensor networks cannot capture pollution variations that occur every 300-500 meters in complex urban environments. The sparse coverage, combined with deployment costs of INR 40 lakhs to 1.6 crores per station, renders comprehensive monitoring economically unfeasible. Image-based air quality estimation emerges as a transformative alternative, yet existing approaches are constrained by limited datasets and poor generalization across cities. This thesis introduces two contributions that establish a new paradigm for scalable air quality monitoring. TRAQID (Traffic-Related Air Quality Image Dataset) provides the first comprehensive multiview dataset with 26,678 synchronized front-rear traffic image pairs, co-located environmental sensors, and systematic coverage across three seasons in Indian urban environments. Benchmark evaluation reveals existing state-of-the-art methods achieve only 75% accuracy on TRAQID, exposing critical limitations in current approaches. AQIFormer, our transformer-based architecture, presents image-based air quality estimation through weather-aware attention mechanisms, adaptive dual-view fusion, and multi-task temporal learning. The architecture achieves 89.96% accuracy on TRAQID—a remarkable 14.96% improvement over existing methods. Most significantly, AQIFormer demonstrates unprecedented cross-city generalization, maintaining 81.67% accuracy on independent Nagpur data with minimal adaptation (5% local samples), exhibiting only 8.29% performance degradation compared to typical 30%+ drops in existing approaches. Comprehensive analysis validates each innovation: dual-view integration contributes 11.46% improvement, multi-task learning enhances discriminability by 11.57%, and attention visualization reveals consistent focus on physically meaningful pollution sources across diverse urban environments. The architecture processes standard traffic imagery in real-time, enabling integration with existing camera infrastructure without specialized hardware. This research establishes the practical foundation for city-scale air quality monitoring at unprecedented spatial resolution and dramatically reduced costs. The robust cross-city performance enables immediate deployment across multiple urban environments, transforming air quality monitoring from expensive, sparse sensor networks to comprehensive, camera-based systems that can protect public health through enhanced environmental awareness.

Abstract

In today’s competitive academic landscape, choosing the right college is a crucial decision for students seeking higher education. Traditional ranking frameworks, such as the National Institutional Ranking Framework (NIRF), provide a comprehensive evaluation of colleges across India based on various parameters, including teaching, learning resources, research output, graduation outcomes, outreach, and institutional perception. While these rankings offer a valuable overview of institutional performance, they may not directly cater to the individual preferences and priorities of students, potentially leading to choices that do not align with personal goals or specific needs. This thesis introduces a novel approach to enhance the utility of NIRF data by developing a userspecific ranking system that tailors college rankings to individual preferences. The proposed system allows users to define their criteria, such as median salary, number of students placed, facilities for Physically Challenged Students (PCS), faculty strength, and research funding, to generate personalized rankings. By integrating these user-defined criteria with NIRF data, the system generates dynamic, customized rankings that adapt to evolving student priorities, providing a more nuanced and relevant assessment of educational institutions. The methodology involves collecting NIRF data and supplementing it with JoSAA cutoff data to provide a comprehensive dataset for analysis. The data is processed using novel techniques, including multi-skyline queries, which expand traditional skyline queries to provide a broader set of ranking options that better reflect user preferences. The system’s architecture, built on the MERN stack (MongoDB, Express.js, React.js, Node.js), supports flexible data handling and user-friendly interaction, allowing for efficient customization and visualization of rankings. By empowering students with personalized insights, this approach facilitates informed decisionmaking and enriches the overall college selection process. It allows students to prioritize aspects of higher education that matter most to them, such as research opportunities, employability, or inclusivity, thus enabling a more tailored approach to selecting a college. This user-specific ranking system is not only a valuable tool for students but also for parents and educational counselors, who can better assist in navigating the complex landscape of higher education based on personalized data. This approach not only supports students in making well-informed decisions but also encourages institutions to strive for excellence in areas that matter most to their prospective students. By enhancing the utility of NIRF data, the proposed system offers a comprehensive and personalized college selection experience, positioning itself as a valuable tool in the evolving landscape of higher education.

Abstract

Water quality monitoring of inland waterbodies is a global challenge due to limited monitoring infrastructure and availability of high-resolution hydrological data. Timely assessment and mitigation of water quality threats is constrained by limited data availability. To address this issue, remote sensing techniques have emerged as a robust tool for continuous real-time monitoring of inland water quality. In tropical regions, there has been a greater emphasis on remote sensing techniques for the continuous monitoring of water quality, mainly in eutrophic or hypereutrophic inland water bodies such as estuaries, rivers, and lakes, while reservoirs have received comparatively less attention. However, worldwide, tropical reservoirs are vulnerable to contamination risks due to their multifunctional roles, i.e., drinking, irrigation, and hydropower generation and long periods of storage times without adequate outflows which are critical for maintaining downstream water quality. Furthermore, the surrounding catchment influence contamination dynamics, as rapid changes in land use/land cover (LULC) can introduce sediments, nutrients, and heavy metals, contributing to elevated levels of physical, biological, and chemical characteristics of water quality parameters. Identifying the critical importance of monitoring tropical reservoirs, the present study focuses on employing remote sensing techniques to estimate water quality parameters, focusing on physio-biological characteristics, mainly in oligotrophic and mesotrophic reservoirs, and underscores the significance of catchment-related factors in influencing contamination levels. This approach is essential for maintaining optimal water quality levels, benefiting both local communities and the surrounding environment. The study selected two tropical reservoirs, Bhadra and Tungabhadra, situated within the same river system and experiencing similar climatic conditions but distinct landscape characteristics. The primarily aim is to develop a modeling framework to map and quantify the spatiotemporal variability of Chlorophyll-a (Chl-a), turbidity and surface water temperature (SWT), along with associated water spread for 2016-2021 using Sentinel 2 and Landsat 8 satellite datasets. The second aim is to examine interconnections among selected water quality parameters, to improve the accuracy of reservoir water quality assessments and identify potential correlations and interdependencies. The third aim is to comprehend how the landscape (land use/land cover (LULC)) and meteorological variables (rainfall, air temperature) of the contributing catchment influence contamination dynamics. The remote sensing-based results, validated with field-based observations measured from the Aquaread AP7000 instrument and secondary data include EOMAP (Earth Observation and Mapping) water quality derived datasets were employed to enhance accuracy and reliability. Chl-a, used as a proxy for nutrient contamination primarily from agricultural runoff, estimated using the Maximum Chlorophyll Index (MCI) derived from Sentinel 2 satellite data in the Google Earth Engine (GEE) platform. The results indicate a consistent and extensive distribution of Chl-a across the entire water surface of the Tungabhadra reservoir, while the Bhadra reservoir exhibited a more limited distribution. During the post-monsoon, increase in Chl-a spread in the Tungabhadra reservoir, primarily attributed to nutrient-rich water inflows from agricultural and urban areas. This notable rise linked to the harvesting of Kharif crops. Additionally, the discharge of untreated sewage further contributes to the degradation of water quality in the Tungabhadra reservoir. In contrast, the Bhadra reservoir, surrounded predominantly by forested areas, maintained water cleanliness and served as a riparian boundary. These findings demonstrate how LULC changes drive nutrient contamination in tropical reservoirs. Turbidity was considered as a proxy for sediments, detected and mapped using Sentinel-2 data. The Normalised Difference Turbidity Index (NDTI) values strongly correlated with EOMAP and field-based observations having values > 10 NTU (nephelometric turbidity unit) with R 2 = 0.81 and standard error estimate (SEE) of 5.3, respectively. In addition, the correlation between red band reflectance values with < 10 NTU, demonstrated a strong relation with an R 2 = 0.74, SEE = 1.7. Combining NDTI and red band reflectance allowed classification into low, moderate, and high turbidity zones. Further analyses indicated that diverse land use patterns, particularly high proportions of urban and agriculture area along with rainfall, significantly influenced the annual and seasonal turbidity variations. Surface water temperature (SWT) is a significant physical parameter affecting aquatic metabolism, nutrient cycling, and contaminant interaction was estimated using Landsat 8 Thermal Infrared (TIR) d

Abstract

Understanding how narrative flow reflects cognitive processes such as memory, schema use, and event organization has long been a goal of cognitive science. Recent advancements in natural language processing, particularly the rise of large language models (LLMs), have introduced new tools for modeling and quantifying linguistic phenomena. Among these, a measure known as sequentiality—proposed by Sap et al. (2022)—aims to capture the narrative coherence of a story using LLM-derived probabilities. However, like many LLM-based approaches, its validity as a cognitively meaningful measure remains underexamined, especially in relation to human-annotated data. This thesis begins with a conceptual replication of the original sequentiality study, using a newly constructed dataset of paired autobiographical and biographical narratives written by the same individuals. Despite theoretical expectations of greater narrative structure in biographies, we fail to replicate the significant differences in sequentiality originally reported. To investigate further, we replicate the original analysis on Sap et al.’s dataset and identify a methodological bias stemming from topic-based components in the formulation. Removing this source of bias reduces the observed difference between narrative types, calling into question the suitability of the original metric for capturing narrative coherence. In the second part of this thesis, we propose a rectified version of the sequentiality measure, excluding the topic term and relying solely on sentence-to-sentence contextual prediction. To empirically validate this variant, we turn to the domain of Automated Essay Scoring (AES), where essays are annotated by human raters across multiple dimensions, including coherence. Using two well-established AES datasets, we demonstrate that the context-only sequentiality score aligns more closely with human judgments of narrative flow than the original formulation. Furthermore, we show that this measure improves the performance of AES systems when incorporated as a feature, highlighting its practical utility as an interpretable and cognitively motivated coherence metric. Together, these findings contribute to the growing literature on using LLMs as models of human linguistic behavior, while emphasizing the necessity for rigorous validation and methodological transparency when deploying such models in cognitive research. This work offers a clearer path toward integrating LLM-derived measures into applied NLP systems and opens avenues for future exploration in interpretability, bias mitigation, and discourse-level evaluation.

Abstract

This thesis presents novel contributions in two primary areas: advancing the efficiency of generative models, particularly normalizing flows, and applying generative models to solve real-world computer vision challenges. In the first part, we introduce significant improvements to normalizing flow architectures through six key innovations: (1) Development of invertible 3 × 3 Convolution layers with mathematically proven necessary and sufficient conditions for invertibility, (2) introduction of a more efficient Quad-coupling layer, (3) Design of a fast and efficient parallel inversion algorithm for k×k convolutional layers, (4) Fast & efficient backpropagation algorithm for inverse of convolution, (5) Using inverse of convolution, in Inverse-Flow, for the forward pass and training it using proposed backpropagation algorithm, and (6) AffineStableSR, a compact and efficient super-resolution model that leverages pre-trained weights and Normalizing Flow layers to reduce parameter count while maintaining performance. These advances maintain model expressiveness while substantially improving computational efficiency compared to existing approaches. The second part demonstrates the practical applications of generative modeling advances across diverse computer vision tasks. This thesis develops: (1) An automated quality assessment system for agricultural produce using Conditional GANs to address class imbalance, data scarcity and annotation challenges, achieving good accuracy in seed purity testing; (2) An unsupervised geological mapping framework utilizing stacked autoencoders for dimensionality reduction, showing improved feature extraction compared to conventional methods; (3) We proposed a privacy preserving method for autonomous driving datasets using on face detection and image inpainting; (4) Utilizing Stable Diffusion based image inpainting for replacing the detected face and license plate to advancing privacy-preserving techniques and ethical considerations in the field.; and (5) An adapted diffusion model for art restoration that effectively handles multiple types of degradation through unified fine-tuning. This thesis advances the theoretical understanding of generative models and their practical applications, demonstrating significant improvements in efficiency, scalability, and real-world utility across multiple domains.

Abstract

The growing availability of biomedical research has increased demand for accessible scientific communication, especially for patients, caregivers, and non-expert readers. However, medical texts, such as journal abstracts, full-length articles, and clinical content, often contain dense information and specialized terminology that can hinder comprehension or cause the ingestion of incomplete or misinterpreted information. This thesis focuses on lay language summarization: the task of transforming complex biomedical paragraphs into simplified, layperson-friendly versions without compromising factual correctness, technical fidelity or essential content. We propose a modular pipeline that leverages domain-specific language models, structured biomedical knowledge, and decision-guided simplification strategies to perform high-quality text simplification in the medical domain. The system is built around three core components: 1. A BioBERT-based term identification module, trained to detect biomedical jargon with high precision. 2. A UMLS-backed definition retriever, which integrates with standard medical ontologies to provide concise lay-level definitions of complex terms. 3. A powerful LLM, which acts as the brain of the operation and implements the reconstruction and iterative rewriting of the content into lay-person friendly summaries. These 3 components were combined into a modular pipeline, capable of simplifying, rephrasing and reconstructing the content, while incorporating the retrieved definitions and maintaining factual consistency. The system is evaluated on a corpus of paragraphs and abstracts drawn from PLOS and eLife journals. Evaluation is conducted using a comprehensive set of metrics across three dimensions: 1. Relevance and Content Preservation, using ROUGE, BLEU, METEOR and BERTScore 2. Readability via FKGL, DCRS, CLI, and LENS. 3. Factual Consistency, measured by AlignScore and SummaC.We also conducted an ablation study to understand the contribution of each component. Additionally, to cover dimensions not captured by the automated evaluation metrics, human feedback was also collected to assess summary quality across readability, clarity, and factuality. The results demonstrated that the proposed pipeline significantly improved readability and user comprehension over current baselines while maintaining a high degree of factual consistency. The modular design allows for transparent interpretation of simplification steps and facilitates future integration with electronic health records, clinical trial explanations, and public health platforms. Overall, this work marks an important step towards accessible medical communication, leveraging the power of domain-specific AI to bridge the knowledge gap between experts and the public.

Abstract

Gynecologic and breast cancers are among the leading causes of morbidity and mortality in women worldwide, posing a significant threat to womens health. It is estimated that roughly one in five women will be diagnosed with cancer during their lifetime, with approximately one in twelve succumbing to the disease. Their molecular and cellular heterogeneity significantly contributes to therapeutic resistance and variability in clinical outcomes. Deciphering these complexities at various resolution is essential for advancing precision oncology. This thesis aims to unravel the molecular landscape of these cancers by leveraging bulk omics, single-cell RNA-seq (scRNA-seq), and spatial transcriptomics datasets. Metabolic reprogramming is a hallmark of cancer and is characterized by significant heterogeneity, presenting challenges for treatment efficacy and patient outcomes. Understanding metabolic adaptation and its heterogeneity in tumor tissues may provide new insights and help in developing personalized therapeutic strategies. We investigated the metabolic alterations of endometrial cancer (EC) to understand the variations in metabolism within tumor samples. Integration of transcriptomics data of EC (RNA-Seq) and the human genome-scale metabolic network was performed to identify the metabolic subtypes of EC and uncover the underlying dysregulated metabolic pathways and reporter metabolites in each subtype. The relationship between metabolic subtypes and clinical variables was explored. Further, we correlated the metabolic changes occurring at the transcriptome level with the genomic alterations. Based on metabolic profile, EC patients were stratified into two subtypes (metabolic subtype-1 and subtype-2) that significantly correlated to patient survival, tumor stages, mutation, and copy number variations. We observed the co-activation of the pentose phosphate pathway, one-carbon metabolism, and genes involved in controlling estrogen levels in metabolic subtype-2, which is linked to poor survival. PNMT and ERBB2 are also upregulated in metabolic subtype-2 samples and present on the same chromosome locus 17q12, which is amplified. PTEN and TP53 mutations show mutually exclusive behavior between subtypes and display a difference in survival. This work identifies metabolic subtypes with distinct characteristics at the transcriptome and genome levels, highlighting the metabolic heterogeneity within EC.

Abstract

Recent efforts to embed ethical reasoning into AI and machine learning systems have focused on training deep learning models to generate moral judgments in response to situational inputs. Parallel developments in the criminal justice system illustrate the application of such technologies, where risk assessment tools are employed to predict defendants’ risk scores. These scores are then used by legal institutions to make decisions related to parole, policing, and sentencing. Together, these approaches reflect a growing trend towards integrating AI into moral decision-making and as a potential replacement for human decision-making. In our work, we critically evaluate moral decision-making in AI systems through a philosophical lens, with particular emphasis on virtue ethics. First, we evaluate Delphi (Jiang et. al., 2021 [1]), which attempts to use the Rawlsian decision procedure as a bottom-up approach to capture abstract moral principles from crowdsourcing moral opinions across a diverse demographic. We critically evaluate this model on the basis of the effectiveness of (a) the current bottom-up approach of incorporating Rawlsian decision procedure, and (b) the proposed hybridization of the former with a top-down approach leading to reflective equilibrium. We argue against (a) on the grounds that the system fails to satisfy the criteria for being a competent moral judge as outlined by Rawls (Rawls, 1971 [2]); further, we argue against (b) on the grounds that reflective equilibrium requires phronesis or practical wisdom, which cannot be embedded into AI systems. We critically evaluate (b) in the light of Irikefe’s neo-Aristotelian, agentbased account of reflective equilibrium (Irikefe, 2019 [3]) to argue against the possibility of AI-ML systems to incorporate phronesis in order to attain reflective equilibrium. We then evaluate COMPAS, a risk assessment software that predicts the risk of recidivism of defendants based on responses to a questionnaire. We investigate COMPAS from a virtue ethics perspective by (1) discussing how punishment in the criminal justice system can be implemented as a practice, as defined within the framework of MacIntyre’s virtue ethics (MacIntyre, 2007 [4]) and (2) analyzing why such a practice cannot be implemented by tools like COMPAS.

Abstract

A control system aims to regulate the behavior of a dynamical system to achieve a desired output by adjusting the input. The control system design specifications for dynamical systems are typically provided in terms of the desired transient response and steady-state response. Performance measures like Rise Time, Settling Time, Steady-State Error, Control Effort, etc., help us understand the effectiveness of the designed controller in real-time. In traditional approaches, based on these specifications, we select design methods such as Root Locus or Bode Plot analysis, and algorithms like ProportionalIntegral-Derivative (PID), Linear Quadratic Regulator (LQR), and Model Predictive Control (MPC). After tuning with the allowed constraints, we may end up with a control system that meets our specifications. However, mathematical models of many real-world systems are not fully known. Meeting design specifications for such unknown dynamical systems is a challenging task. Partial or approximate system models often lead to inaccuracies due to internal modeling errors that are not always avoidable or correctable. Therefore, there is a need to explore control strategies that do not depend on mathematical models. Model-free Reinforcement Learning (RL) methods have demonstrated strong performance in this context. Their performance was comparable to traditional control design methods with system knowledge, and sometimes even better. This is due to their ability to learn optimal policies through interaction with the environment and adaptability to changes in real-time. Techniques such as Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Twin Delay Deep Deterministic Policy Gradient (TD3) have shown significant potential in handling complex control tasks without explicit system models. This thesis presents a new approach for designing controllers for dynamical systems with unknown mathematical models using a model-free RL framework based on the TD3 algorithm. The proposed method enables an RL agent to learn and shape the error dynamics and overall system response in trajectory tracking, particularly in state regulation tasks, through a suitably designed reward function. The effectiveness of the approach is validated on first-order and second-order linear systems, as well as a nonlinear pendulum system. The resulting controller significantly influences both the transient and steady-state behavior of an unknown system while adhering to control and input constraints. By systematically modifying the reward parameters, we analyze how reward shaping during training affects test-time performance metrics. This analysis establishes the relationship between reward design and control performance, providing a means to tune reward parameters to meet desired specifications. Furthermore, we introduce a data-driven optimization framework using Particle Swarm Optimization (PSO) to identify sets of reward parameters that satisfy given performance requirements. The method achieves exponential convergence of tracking error dynamics in linear systems without requiring explicit knowledge of system models.