Abstract

Human intelligence relies on compositional generalization: the ability to interpret novel situations by flexibly combining familiar concepts and relational structures. This thesis investigates compositionality in vision-language models (VLMs), focusing on their ability to understand and generalise across visual (images, videos) and linguistic inputs. In the first part, we introduce VELOCITI, a benchmark for evaluating compositional understanding in video-language models through a suite of entailment tasks. Unlike prior compositionality benchmarks constrained to single-agent videos, VELOCITI captures the complexity of real-world videos involving multiple agents and dynamic interactions. VELOCITI assesses how well models recognize and bind agents, actions, and temporal events using both text-inspired and in-video counterfactual negations. In the second part, we probe the internal activations of VLMs to understand how concepts in an image are bound to their attributes and references in text. Extending the Binding ID mechanism in language models, we demonstrate that VLMs construct binding ID vectors in the activations of both image tokens and their textual references, enabling in-context concept association. Together, these contributions advance our understanding of compositional reasoning in VLMs and offer tools for probing their capabilities.

Abstract

This thesis examines the utilisation of hybrid methodologies, integrating traditional textual analysis with computational techniques, to optimise research efficiency in the analysis of political discourse. This endeavour aligns with the ongoing paradigm shift within the Digital Humanities, driven by advancements in computational and machine learning technologies. Specifically, two distinct projects analyse the public addresses of the incumbent Prime Minister, Narendra Modi. The first project investigates Mann ki Baat, a series of radio broadcasts initiated during the first Modi administration. This study aims to ascertain the primary objective of the series. The second project analyses the Prime Minister’s Independence Day speeches (15th August), exploring the contrasts in narrative style and content between the two terms of the current administration and the rationale behind these perceived changes. The first study examining the Mann ki Baat radio broadcasts utilises a suite of Natural Language Processing (NLP) techniques to generate initial observations. Subsequently, classical textual analysis is applied to a carefully curated corpus of 25 speeches, selected for their temporal alignment with significant socio-political events in India, to validate, substantiate, and extend these findings. The research demonstrates that the series functions as an aesthetic imperative, representing a large-scale governmental initiative that uses state machinery and financial resources to propagate the values and ideals of the BJP as a political entity over the overarching NDA government. The second study divides Narendra Modi’s ten Indian Independence Day addresses according to his terms of office. To facilitate rigorous analysis, a comprehensive annotation procedure was implemented, wherein the thematic content of each sentence was systematically categorised. Subsequently, a comparative analysis was performed, contrasting the content and narrative styles of the two sets of speeches. The findings reveal a significant thematic transition: from an emphasis on inclusivity during the first term to a focus on infrastructural development and national self-sufficiency in the second. Furthermore, discernible shifts in narrative style were observed, including a consistent augmentation of monoreligious allusions and direct appeals to a gendered audience. By integrating the conclusions derived from both projects, albeit obtained through different methodologies, a coherent understanding of the overarching narrative that the government hasarticulated during its tenure in power can be established. Furthermore, this research demonstrates the feasibility of employing hybrid methodologies in the analysis of political rhetoric.

Abstract

This thesis addresses the problem of achieving fairness in probabilistic binary classification in the presence of binary protected groups. In many practical scenarios, classifiers assign scores, and decision thresholds are applied post hoc based on the desired trade-off between false positives and false negatives. However, the use of a fixed classifier with arbitrary thresholds may result in disparate treatment across protected groups. To ensure equitable outcomes, we introduce a fairness criterion, denoted εp-Equalized ROC, which requires that the Lp norm between the false positive rates (FPRs) and true positive rates (TPRs) of the protected groups remains within a predefined bound ε—regardless of the threshold chosen. To operationalize this notion of fairness, we propose a post-processing approach that transforms the outputs of an existing (potentially unfair) classifier into a randomized and threshold-agnostic classifier that satisfies ε1-Equalized ROC. Central to this approach is a novel threshold query model over ROC curves, which enables controlled manipulation of the classifier’s behavior across groups. We characterize the inherent trade-off between fairness and utility by deriving a theoretical lower bound on AUC loss and proving that some AUC reduction is inevitable under fairness constraints. To realize our fairness objective in practice, we develop a linear-time algorithm, FROC, which guarantees ε1-Equalized ROCcompliance. We prove that under reasonable assumptions, FROC achieves optimality with respect to the minimal necessary AUC degradation. Extensive experiments on real-world datasets—using classifiers such as Weighted Ensemble L2, Random Forest (Gini), and FNNC—demonstrate that FROC is both theoretically sound and empirically effective across multiple fairness and performance metrics. The thesis is structured into six chapters, beginning with an introduction to algorithmic bias and culminating in theoretical analysis, algorithmic development, and empirical validation. Through this work, we contribute a principled and practical framework for ensuring fairness in score-based classification systems

Abstract

The need for photorealistic head avatars has risen in the past decades, owing to the rising interest in the AR/VR media formats. An accurate representation of the head will be required in the near future, which is essential to facilitate communication between users, essentially enabling telepresence. The need for an improved in-person form of remote communication was made more clear during the recent COVID-19 pandemic and the ensuing lockdown, where millions of people had to stay away from their families and workplace for an extensive period of time. Besides, realistic facial avatars have proved immensely helpful in the movie and gaming industry, where they have often been used to either modify the actors’ appearance itself, or to drive an entirely virtual but realistically looking digital character, depending on the demands of the narrative. Capturing and reconstructing a realistic-looking head-avatar is not trivial, and requires an expensive setup of multiple synced cameras and lights, and a mathematical understanding of how light interacts with the skin, hair, cornea, etc. The capture of each subject is laborious and time-consuming. The creation of digital faces that are indistinguishable from real ones is a formidable challenge due to the ”uncanny valley” phenomenon, where even minor deviations from realistic appearance can render a digital face unsettling to human observers. However, to achieve the applications of realistic head avatars in telepresence and AR/VR, the capture and thus creation of realistic digital replicas must be made accessible. Therefore, a need has arisen to search for methods that can reconstruct and create digital replicas that are photorealistic but also cheap. Our thesis aims to tackle this problem statement in two ways, one from the perspective of digital replica generation and the other from the perspective of creating a digital replica through reconstruction. Our initial approach to make the creation of digital replicas efficient is a textured head generation method conditioned on a descriptive text. We aim to create a method that can generate a realistic-looking head avatar from a text description in an efficient manner, without requiring the manual intervention of artists or the use of highly specialised software like Blender or Maya. Therefore, it can generate textured head assets within seconds. However, the texture-based synthesis approach suffered from reduced realism because of the effects of baked-in lighting. Therefore, an approach is required that could construct a head avatar along with accurate material properties, such that it can be placed in any environment. Our following work proposes a method to reconstruct a relightable head avatar using just a smartphone. Existing works that focus on creating a digital replica from a smartphone do not create a relightable head avatar or require a large amount of data prior to generating a head avatar with appropriate material properties. Our proposed method introduces a novel capture strategy that can act as a smallscale replacement of lightstage-based capture, while only utilising a smartphone. We also propose an efficient head representation which allows real-time rendering and relighting, making it suitable for the application of telepresence. We evaluate our proposed methods on public datasets, and conclude that our methods outperform state-of-the-art methods, both in the realms of text-based generation and capture-based reconstruction. We also release our own dataset, recorded using our novel capture setup. Additionally, we discuss our methods’ limitations, offering insights into potential improvements and outlining promising future research directions. We hope that this thesis will substantially impact the field and accelerate progress in 3D Head generation and reconstruction.

Abstract

he escalating prevalence of Oral Potentially Malignant Disorders (OPMDs) and oral cancer in low- and middle-income countries presents a critical challenge, exacerbated by limited resources that hin- der population screening in remote areas. The study evaluates the efficacy of artificial intelligence (AI) and digital imaging diagnostics as tools for OPMD detection in the Indian population, utilizing smartphone-captured oral cavity images. Trained front-line healthcare workers (FHWs) contributed a dataset comprising 1,120 suspicious and 1,058 non-suspicious images. Various deep-learning mod- els, including DenseNets and Swin Transformers, were assessed for image-classification performance. The best-performing model was then tested on an independent external set of 440 images collected by untrained FHWs. DenseNet201 and Swin Transformer (base) models exhibited high classification accuracy on the internal test set, achieving F1-scores of 0.84 (CI 0.79–0.89) and 0.83 (CI 0.78–0.88). However, performance declined on the external set—characterized by significant variation in image quality—with DenseNet201 yielding the highest F1-score of 0.73 (CI 0.67–0.78). The AI model demon- strates potential for identifying suspicious versus non-suspicious oral lesions via photographic images. This image-based solution holds promise for facilitating early screening, detection, and timely referral for OPMDs.

Abstract

The digital age has amplified the importance of inclusive and comprehensive information dissemination, yet many low-resource languages remain underrepresented in this sphere. The lack of advanced natural language processing (NLP) tools for these languages hampers the ability to effectively combat misinformation, detect abusive content, and provide enriched informational resources. Addressing these challenges is crucial for fostering digital inclusivity and ensuring that speakers of low-resource languages have equitable access to reliable and enriched information. Our first study focuses on detecting fake news in Dravidian languages by leveraging Google’s MuRIL (Multilingual Representations for Indian Languages) model. By fine-tuning the ”MuRIL-base-cased” variant with a carefully curated dataset of labeled comments and posts in Dravidian languages, we enhance the model’s ability to differentiate between authentic and deceptive content. The performance evaluation using standard metrics demonstrates the effectiveness of MuRIL in maintaining a safer digital ecosystem for speakers of Dravidian languages. In the second study, we address the pressing need for detecting abusive language in multimodal content, particularly in Tamil videos, and performing sentiment analysis for Tamil and Malayalam videos. Our novel framework integrates state-of-the-art models such as Multiscale Vision Transformers (MViT) for video analysis, OpenL3 for audio analysis, and bert-base-multilingual-cased for textual analysis. By combining these multimodal features, our approach significantly improves the accuracy of abusive content detection and sentiment categorization, contributing to a safer and more inclusive online environment for Dravidian language speakers. The third study tackles the challenge of information scarcity in the very low-resource language of Mizo. We investigate a methodology to generate holistic summaries of Mizo news articles, enriched with information from English-language news to enhance coverage. By providing a dataset of 500 Mizo news articles with corresponding enriched summaries, our approach significantly improves the comprehensiveness and accessibility of news content for Mizo speakers. Human evaluations confirm the effectiveness of our method in meeting the information needs of this community. Together, these studies highlight the critical importance of developing advanced NLP techniques for underrepresented languages. By addressing fake news detection, abusive language identification, and information enrichment in Dravidian and Mizo languages, our research makes significant strides toward digital inclusivity and safety, ultimately enhancing the digital experience for speakers of these languages and paving the way for future innovations in multilingual and multimodal content analysis.

Abstract

Environmental law in India has played a pivotal role in balancing economic development with ecological sustainability, with the Supreme Court emerging as a crucial actor in shaping environmental jurisprudence. However, the ways in which courts frame environmental concerns significantly influence judicial decision-making, policy formulation, and public discourse. This thesis investigates the legal narratives surrounding environmental issues in Indian judicial decisions, analyzing how courts construct arguments around environmental protection, economic growth, and social justice. By applying computational methods to a curated dataset of Supreme Court judgments, this research uncovers patterns in legal argumentation and broader ideological shifts within environmental jurisprudence. While traditional legal scholarship has extensively examined the doctrinal foundations of environmental law, there remains a significant gap in understanding the rhetorical and structural aspects of judicial decision-making. This study addresses this gap by adopting a computational humanities approach, employing natural language processing (NLP) and statistical analysis to identify dominant themes and argumentative strategies in environmental case law. Through topic modeling and citation analysis, this research maps the evolution of environmental jurisprudence in India, highlighting the judiciary’s role in shaping governance mechanisms and regulatory frameworks. The findings reveal critical insights into the judiciary’s framing of environmental issues. The study identifies the predominance of individual appellants in environmental litigation, suggesting a growing public engagement in environmental advocacy, but also underscores the limited participation of institutional actors such as state governments and NGOs. The thematic analysis of case topics shows that courts focus heavily on constitutional rights, pollution control, and land disputes, while areas such as biodiversity conservation and hazardous waste management receive comparatively less judicial attention. Additionally, the temporal analysis of forest rights cases demonstrates how legislative interventions, such as the Forest Rights Act (2006), directly influence litigation trends, reaffirming the judiciary’s role in shaping environmental governance. This study also examines the legal principles that form the foundation of environmental jurisprudence in India, particularly the interpretations of Articles 14, 21, and 32 of the Indian Constitution. By analyzing landmark cases, the research highlights how the Supreme Court has expanded environmental rights under Article 21’s ”Right to Life,” employed the Public Trust Doctrine, and developed principles such as ”Polluter Pays” and ”Precautionary Principle.” The increasing reliance on Public Interest Litigations (PILs) has democratized access to environmental justice but has also led to concerns about the potential misuse of this legal mechanism. The study’s methodological innovation lies in leveraging computational techniques for legal analysis. By integrating large-scale text mining with qualitative legal interpretation, this thesis offers a data-driven perspective on judicial reasoning. The citation analysis of frequently referenced cases, laws, and precedents provides insights into how legal arguments evolve over time. The application of NLP techniques enables the identification of implicit biases in judicial narratives, revealing recurring themes and the ideological underpinnings of environmental litigation. Despite its contributions, the research acknowledges certain limitations, including the reliance on a specific dataset and the challenges of drawing strong correlations between case topics and appellant types. Future research could expand this work by conducting comparative studies across different jurisdictions, exploring regional variations in environmental litigation, and incorporating qualitative insights from legal practitioners. Ultimately, this thesis contributes to a deeper understanding of how environmental concerns are framed, contested, and adjudicated in India’s judicial system. By bridging environmental law with computational humanities, this study not only enriches legal scholarship but also provides valuable insights for policymakers, environmental activists, and legal practitioners. As climate change and sustainability debates gain prominence, the role of the judiciary in environmental governance will continue to shape the trajectory of India’s environmental policies and legal frameworks

Abstract

The exponential growth of big data, driven by rapid technological advancements, has transformed the landscape of data analysis and business decision-making. The integration of data from the Internet of Things (IoT) devices, healthcare, manufacturing, banking, and other applications has opened up new avenues for deriving valuable insights and enhancing business operations. Organizations leverage powerful analytics to drive strategic decisions, identify opportunities, and improve performance. However, the vast increase in data usage and complexity brings significant security and privacy challenges. Big data relies heavily on infrastructure, cloud, and edge computing, all of which pose inherent security risks. Untrusted applications and insufficient security measures can introduce vulnerabilities into enterprise systems. Additionally, the widespread use of mobile and IoT devices increases the risk of unauthorized access and misuse of sensitive data. Enterprises must prioritize defining secure access privileges and implementing robust security protocols to prevent potential data misuse. As data sharing and source authentication become more prevalent, trust between parties remains a critical issue. Governments, for instance, are cautious about sharing data due to national security concerns, while individuals worry about the privacy of their personal information. The involvement of multiple stakeholders often leads to poor control over data transfer, hindering the development of the big data industry. The presence of personally identifiable information (PII), trade secrets, and other sensitive details raises significant privacy concerns. Balancing the utility of data analysis with the protection of individual privacy requires employing various cryptographic, anonymization, and access control methods. Compliance with privacy regulations, such as the “General Data Protection Regulation (GDPR) in the European Union”, is crucial for organizations handling big data. Non-compliance can lead to severe legal and financial consequences. Therefore, secure big data analytics requires secure data transfer and privacy-preserving that necessitates a combination of technical, organizational, and legal measures to ensure that individual privacy is protected while still enabling the derivation of valuable insights from large-scale data analysis. Thus, the ongoing challenge of secure big data analytics demands innovative solutions that can effectively secure data with privacy protection. Existing methods often struggle with complexity, scalability, and the ability to ensure secure big data analytics while maintaining privacy and confidentiality of data. Organizations must continuously adapt and enhance their strategies to keep pace with the evolving landscape of big data and its associated risks. The first study in this thesis presents a comprehensive approach that utilizes data of smart devices in IoT. This data is crucial for building intelligent applications using machine learning.Typically, data from these devices is collected into data centers for training machine learning models. The traditional centralized training approach, which requires transferring data from devices to a central server, is inefficient due to privacy concerns and bandwidth limitations, as users are often reluctant to share their data with centralized data centers. To address these issues, we propose an efficient hybrid secure federated learning approach integrated with blockchain technology. This method allows models to be securely trained locally on devices, with the model and its parameters stored in the blockchain for traceability and immutability. A detailed security and performance analysis demonstrates the efficacy of this proposed approach, highlighting its security, resilience against numerous attacks, and cost-effectiveness in computation and communication compared to existing schemes. We introduce a secured federated learning approach for big data analytics based on an authenticated key agreement protocol and blockchain. We utilize the “Genetic Algorithm Neural Network (GANN)” to train our model and propose a secure and efficient authentication and key agreement scheme to ensure the secure exchange of model parameters in federated learning. The transactions created as model parameters among IoT smart devices are securely stored in blocks and added to the blockchain using the “Practical Byzantine Fault Tolerance (PBFT)” consensus protocol. Experiments and performance analysis of our proposed machine learning model indicate its efficiency under various attack scenarios. Additionally, a detailed formal security analysis and verification using the widely-used Scyther simulation tool show that the proposed scheme is resilient against several attacks common in the IoT environment. A rigorous comparative study confirms that the proposed scheme offers an accurate, secure, and resilient federated ma

Abstract

Experiments at the Large Hadron Collider (LHC) aim to discover exotic particles by colliding protons at extremely high speeds and analyzing the resulting data. A key challenge in these processes is distinguishing between experimental observations and theoretical predictions, particularly in separating signals of interest from background noise. Traditionally, cut-based methods have been employed for this purpose; however, while effective to some extent, these methods are often inefficient for optimal signal-background classification. In this thesis, we explore various machine learning-based approaches to improve signal-background classification, leveraging their potential to outperform traditional techniques. Additionally, we investigate a phenomenological search for a Z ′ boson decaying into right-handed neutrinos, utilizing these advanced methods to achieve significantly better classification accuracy compared to earlier cut-based analyses. This work highlights the transformative role of machine learning in enhancing the sensitivity and efficiency of particle physics research.

Abstract

Toxicity in online gaming is more than just an unfortunate side effect of competitive play; it is a deeply embedded issue shaped by game design, player dynamics, and broader social structures. While research on gaming toxicity has largely centred on Western contexts, this study shifts focus to India, where unique cultural, linguistic, and regional factors influence how toxicity is experienced and perpetuated. Gaming in India is a rapidly growing space, yet it remains deeply exclusionary in many ways, with toxicity often targeting players based on gender, language, region, or religious identity. Through an ethnographic approach, including in-depth interviews and observational co-playing, this study investigates how toxicity manifests, how players rationalise or resist it, and how online interactions reflect broader societal divisions. One of the key concerns explored in this research is the role of game design and matchmaking systems in reinforcing bias and exclusion. Most matchmaking algorithms prioritise skill-based ranking and latency optimisation but fail to consider the social dimensions of player experiences. This oversight results in structures that inadvertently foster hostility, as players are matched without consideration for linguistic compatibility, cultural norms, or pre-existing biases. The study reveals that players from marginalised groups, such as women, non-binary players, and linguistic minorities, frequently encounter discrimination and harassment, often leading them to modify their behaviour, hide aspects of their identity, or disengage from competitive gaming entirely. The findings also highlight the normalisation of toxicity in online spaces, where aggressive behaviour is often framed as a natural or even necessary part of competition. In Indian gaming communities, toxicity is deeply gendered, with female players regularly facing scepticism, harassment, and pressure to prove their competence. Additionally, linguistic and regional biases shape in-game interactions, where players who speak certain languages or exhibit accents associated with particular regions are frequently targeted. The study also examines how asymmetric and symmetric game structures influence the dynamics of toxicity, with different formats fostering distinct patterns of aggression and exclusion. Beyond player interactions, this study considers the impact of external influences, including streamers, gaming influencers, and community echo chambers, in legitimising or amplifying toxic behaviour. Popular streamers often set behavioural norms for younger audiences, reinforcing toxic patterns that trickle down into everyday gaming culture. While some gaming personalities attempt to promote inclusivity, others actively foster hostility, normalising aggression and harassment in ways that extend beyond individual matches or specific game titles.Rather than relying solely on reactive measures like reporting and banning, there is a need to integrate proactive design changes, such as improved matchmaking filters, greater support for linguistic diversity, and better moderation tools, that take into account the cultural and social realities of gaming in the Global South. This study is also grounded in a postcolonial perspective, recognising that global gaming platforms often reflect Western norms, values, and modes of communication that do not always align with the lived experiences of players in regions like India. The dominance of English, the lack of region-sensitive moderation, and the design of social systems in games often position Indian players at the margins, forcing them to adapt to structures that were never made with their context in mind. Toxicity, in this sense, is not only a behavioural issue but also a symptom of deeper structural inequalities rooted in global digital hierarchies. By situating Indian players within this broader postcolonial framework, this research highlights how digital spaces continue to reproduce older power dynamics, demanding more localised and culturally sensitive approaches to gaming design and governance. This study contributes to critical discussions on gaming toxicity, algorithmic bias, and social disparities in digital environments, offering valuable insights for researchers, game developers, and policymakers. As online gaming continues to grow in India and other non-Western contexts, understanding how toxicity operates in these spaces is crucial for fostering more inclusive and equitable multiplayer ecosystems.