Abstract

Traffic flow prediction is a critical component of intelligent transportation systems, with direct implications for congestion mitigation, safety enhancement, and efficient travel planning. Given the complex spatio-temporal nature of traffic data, a wide range of hybrid deep learning models—such as CNN-LSTM, ConvLSTM, and Temporal Convolutional Networks (TCNs)— have been explored to capture spatial, temporal, and periodic dependencies effectively. In this work, we perform a comprehensive comparison of deep learning models with and without periodicity to evaluate their prediction accuracies. To address the challenge of hyperparameter tuning in these models, we propose a Genetic Algorithm (GA)-based optimization framework for CNN-LSTM, ConvLSTM, and a novel GA-TCN architecture, all of which demonstrate improved performance on benchmark datasets. Furthermore, we introduce a purely temporal deep learning model named Grid LSTM-based Attention Modelling for Traffic Flow Prediction (GLSTM-A), which leverages a combination of Grid LSTM for long-term dependencies, a standard LSTM for recent trends, and a custom attention mechanism to automatically prioritize significant temporal features. GLSTM-A exhibits superior prediction accuracy and memory efficiency compared to existing temporal models such as TCN, LSTM, and Bi-LSTM. Finally, to tackle the challenge posed by limited long-term traffic datasets, we develop TrafficFlowGAN—a GAN-based time-series data generation framework that effectively captures temporal patterns through a joint supervised and adversarial learning process. The synthetic data generated by TrafficFlowGAN closely resembles real-world traffic patterns, thereby enhancing the robustness and accuracy of downstream prediction models. Extensive experimental evaluations and ablation studies validate the efficiency of the proposed models across various traffic forecasting scenarios.

Abstract

The rapid adoption of Large Language Models (LLMs) as tools for annotation and evaluation tasks in Natural Language Processing (NLP) has led to important questions about their reliability, robustness, and alignment with human expectations. This thesis investigates these concerns across three connected lines of inquiry: • the indistinguishability of LLM-generated annotations from human annotations, • human alignment and reliability of LLMs as subjective judges of language quality, and • the susceptibility of LLM-judges to misinformation attacks in evaluation settings. In the first part of this thesis, we study whether annotations made by LLMs can be reliably distinguished from those created by humans. Previous research has claimed that LLMs behave as ’human-like annotators’, motivating our work to rigorously test this hypothesis. We frame this as a classification task: Given a dataset of annotations, can a model detect their origin? Surprisingly, our findings indicate that even state-of-the-art classifiers achieve near-chance performance, with accuracy not exceeding 51%. These results offer strong empirical evidence that LLMs produce annotations that are indistinguishable from human annotators in most settings, validating the ’human-like’ claim from a discriminative modeling point of view. Building on this, we explore whether LLMs can function as reliable evaluators: a more nuanced role that goes beyond labeling and involves subjective scoring of text quality (e.g., summaries, open-ended responses). In this phase, we test how well LLM-generated judgments align with human evaluations and how robust they are to subtle changes in prompt phrasing. We introduce controlled prompt perturbations and evaluate the consistency of LLM scores and textual justifications. The results show a significant lack of robustness: Small, often misleading perturbations lead to large changes in judgment outputs. These inconsistencies expose limitations in using LLMs as trustworthy evaluators, especially in high-stakes or subjective settings. The final part of this thesis presents a preliminary exploration into the robustness of LLM-based evaluators when exposed to misinformation, marking a conceptual shift from controlled perturbations to real-world-inspired adversarial scenarios. We introduce a novel framework that systematically injects misinformation into prompts, grounded in a taxonomy of ten misinformation types. This study probes how these manipulations affect LLM judgments, analyzing both score alignment and justification consistency. Contrary to our initial expectations, the results suggest that LLM-judges demonstrate a surprising degree of resilience. Scores and justifications remain largely consistent across many misinformation types, even when factual content is altered. However, this robustness appears to be superficial, with the LLM-Judge often failing to explicitly recognize or reflect awareness of being misled in its justifications. Rather than detecting and correcting misinformation, they tend to preserve their prior judgments, raising concerns about unconscious robustness without epistemic awareness. These early findings open new avenues for future research. Comparative assessment, misinformation detection, and deeper semantic analysis may reveal subtler vulnerabilities. More broadly, this work lays the groundwork for building misinformation-aware evaluation systems and motivates the development of LLMs that not only evaluate well but do so with fact-sensitive reasoning. Collectively, this thesis explores the evolving role of LLMs as annotation and evaluation agents. While they show promise as ’human-like’ annotators, their behavior as evaluators under adversarial and misinformation-rich contexts highlights both strengths and blind spots. By offering early insights into their robustness and limitations, we provide a foundation for future work aimed at building trustworthy, interpretable, and context-aware LLM judges.

Abstract

Large Language Models (LLMs) are powerful AI tools with broad abilities, but they still struggle with tasks that require deep understanding and complex, non-standard reasoning. This thesis examines two major challenges: how well LLMs understand and interpret cross-cultural communication, and how to improve their advanced problem-solving skills like lateral thinking and context-aware support. It evaluates how well current LLMs handle these tasks and explores two main strategies to improve their performance: advanced prompting and reinforcement learning-based fine-tuning. The first part of the research focuses on cultural understanding. A user study of book reviews shows that most (83%) contain Culture-Specific Items (CSIs), often making cross-cultural understanding difficult. We test different LLMs—including GPT-4o and smaller open models like Aya and Gemma—on how well they can detect and categorize these CSIs. The results show mixed strengths across models (e.g., Gemma-2 does well with social references, Aya with customs), but all models show a noticeable Western bias compared to human evaluations. To support future work, we also provide standardized datasets. The second part explores LLM performance in complex reasoning tasks such as puzzles that require creative and unconventional thinking. We find that advanced prompting—using both static and dynamic examples, along with model-generated reasoning steps—greatly improves results with Gemini 1.0 Pro over basic methods. However, models still do not reach human-level performance, suggesting that prompting is a promising way to improve complex reasoning, even if it is not yet perfect. Finally, the thesis focuses on generating context-aware assistance for solving challenging math problems, particularly those at the Olympiad level. We propose a framework for producing diverse, multistage synthetic hints tailored to different student needs. These hints cover a range of instructional types, including: beginning steps to help students start; relevant theorems and definitions to build conceptual grounding; equivalent examples for analogical learning; next steps and summaries of remaining steps based on student progress; identification of unused information from the problem; and correctionoriented hints to help students recognize and fix mistakes. Using this framework, we fine-tuned Small Language Models (SLMs) through several reinforcement learning approaches—Supervised Fine-Tuning (SFT), Kahneman-Tversky Optimization (KTO), and Odds Ratio Policy Optimization (ORPO)—to train models to generate helpful, targeted mathematical hints. This not only addresses a key gap in automated pedagogical support for difficult problems but also tackles an important research question: how effectively can these reinforcement learning methods align a smaller model with significantly fewer parameters to synthetic data generated by a much larger LLM? Overall, we work on multi-dimensional evaluation of LLMs in cultural understanding and advanced problem-solving. It highlights current limitations, reveals existing biases, and demonstrates that prompting and fine-tuning together can build more capable, adaptive, and culturally aware AI systems.

Abstract

Re-identification (ReID) is a critical challenge in computer vision, predominantly studied in the context of pedestrians and vehicles. However, robust object-instance ReID, which has significant implications for tasks such as autonomous exploration, long-term perception, and scene understanding, remains underexplored. In this work, we address this gap by proposing a novel dual-path object-instance re-identification transformer architecture that integrates multimodal RGB and depth information. By leveraging depth data, we demonstrate improvements in ReID across scenes that are cluttered or have varying illumination conditions. Additionally, we develop a ReID-based localization framework that enables accurate camera localization and pose identification across different viewpoints. We validate our methods using two custom-built RGB-D datasets, as well as multiple sequences from the open-source TUM RGB-D datasets. Our approach demonstrates significant improvements in both object instance ReID (mAP of 75.18) and localization accuracy (success rate of 83% on TUM-RGBD), highlighting the essential role of object ReID in advancing robotic perception. Our models, frameworks, and datasets have been made publicly available.

Abstract

This thesis investigates two aspects of how humans engage with visual content: how it is described and how it is remembered. The first part focusses on Audio Descriptions (ADs), which convey essential on-screen information, allowing visually impaired audiences to follow videos. To be effective, ADs must form a coherent sequence that helps listeners visualise the unfolding scene, rather than describing isolated moments. However, most automatic methods generate each AD independently, often resulting in repetitive, incoherent descriptions. To address this, we propose a training-free method, CoherentAD, that first generates multiple candidate descriptions for each AD time interval, and then performs auto-regressive selection across the sequence to form a coherent and informative narrative. To evaluate AD sequences holistically, we introduce a sequence-level metric, StoryRecall, which measures how well the predicted ADs convey the ground truth narrative, alongside repetition metrics that capture the redundancy across consecutive AD outputs. Our method produces coherent AD sequences with enhanced narrative understanding, outperforming prior approaches that rely on independent generations. The second part investigates Video Memorability. Understanding what makes a video memorable has important applications in advertising and education technology. Towards this goal, we investigate spatio-temporal attention mechanisms underlying video memorability. Different from previous works that fuse multiple features, we adopt a simple CNN+Transformer architecture that enables analysis of spatio-temporal attention while matching state-of-the-art (SoTA) performance on video memorability prediction. We compare model attention against human gaze fixations collected through a small-scale eye-tracking study where humans perform the video memory task. We uncover the following insights: (i) Quantitative saliency metrics show that our model, trained only to predict a memorability score, exhibits similar spatial attention patterns to human gaze, especially for more memorable videos. (ii) The model assigns greater importance to initial frames in a video, mimicking human attention patterns. (iii) Panoptic segmentation reveals that both (model and humans) assign a greater share of attention to things and less attention to stuff as compared to their occurrence probability

Abstract

This thesis investigates how entities are shaped by the discourse in which they appear. Rather than being isolated mentions, entities take on meaning through their interactions with the surrounding context, both in terms of reference and interpretation: the identity of an entity can shift depending on how it is invoked, and the narrative framing around it can significantly influence how it is perceived by a reader. At its core, the research addresses these two complementary dimensions of entity analysis: the challenge of tracking entity references across text (coreference resolution) and the interpretive challenge of understanding how entities are characterized through their framing within the text (entity role classification). While these tasks represent distinct technical problems, they are unified by their contribution to richer discourse comprehension and their shared goal of moving NLP systems beyond surface-level language processing toward deeper semantic understanding. The first part of this research focuses on coreference resolution for South Asian languages, addressing a significant gap in current NLP research where these morphologically rich and syntactically diverse languages remain underrepresented. To overcome the lack of suitable evaluation resources, we develop mGAP, a novel dataset containing coreference annotations across 27 South Asian languages, created through a dedicated translation and alignment pipeline from English reference data. This dataset provided the foundation for investigating cross-lingual transfer learning strategies designed to address the low-resource nature of many of these languages. Through systematic experimentation, we examine which languages serve as effective source languages for training when target language data is scarce, offering insights into making use of linguistic similarities for enhanced cross-lingual adaptation. The second part shifts focus to entity framing in news media, where the research question evolves from identifying what an entity refers to, to understanding what narrative role that entity plays within the text. The framing of the text is pivotal in how a reader would perceive the entity being described. For this task, we developed a two-stage pipeline for classifying entities into roles such as protagonist, antagonist, and innocent, as well as more fine-grained narrative categories. The pipeline first employs entity-centric summarization to extract relevant contextual information surrounding target entities, followed by classification using a DeBERTa v3-based model enhanced with contrastive learning techniques. This approach addresses the significant challenge of determining how entities are portrayed based on surrounding textual cues, contributing further to computational methods for analyzing media bias and narrative construction.

Abstract

India is experiencing an innovation upswing in many fields such as: pharmaceuticals, agriculture, space technology and especially in the area of IT/computer science. The Indian IT industry has a lot to offer: lower human resource costs, large talent pool, favourable government policies, access to capital, and advanced infrastructure that makes India the cost-effective Silicon Valley of the world. It also hosts a significant number of world-class research institutions producing handsome volumes of deep-tech computer science research. However, it has been observed that much of this research is unable to productise and become usable for markets in the real world This has resulted in a high volume of research with potential to disrupt markets but fail to carry through the ‘lab to land’ journey for a research output. This thesis is a study to understand the translation process of a research as a marketable product and to identify factors that impede or enable this translation. The productisation window for any relevant computer science research work is largely time sensitive. [1] Recent explosion in AI adoption fueled with a digital infrastructure has made many deep-tech Computer Science research areas to get production ready very fast. It has an associated shelf life during which it should be explored and leveraged for any sort of translational or productisation activity. Therefore, as a community engaging closely with research problems bridging disconnections between research labs and industry/market requires critical and timely interventions. While there are many research translation models available as a concept or practice, many are ineffective in explaining a seamless / smooth transition of any research work into an industry-consumable technology or product. Based on tacit knowledge & existing models of translation it is apparent to observe that due to the nature of the research ecosystem, its processes and the associated stakeholders within it, identifying potential translation flow is far from trivial. In order to understand the entire journey of a potential research translation, this thesis proposes a productisation flow that considers all potential channels through which a research work in computer science can transition into a market relevant product or an industry consumable technology. This thesis will cover the productisation flow in detail in the later chapters. In this study, we focus on understanding this gap in productisation from the perspective of research scholars who actually work on these projects. After an extensive literature study, the thesis focuses on highlighting the gap in considering ‘researchers’ as an active stakeholder in this entire productisation process. As a result, this thesis proposes a productisation flow of deep-tech research from lab to market translations that highlights the participation of researchers across various translational channels and their stages. This qualifies the need to understand reasons why research flow from lab to land is enabled or inhibited wrt researchers.

Abstract

This study critically analyses portrayals of masculinity on Over The Top (OTT) platforms in India, which have seen a marked rise in viewership since the beginning of the COVID-19 pandemic. Netflix, currently the largest investor in original content in India, offers a range of curated material in 31 languages. Masculinity over the years in Indian cinema, particularly the Bollywood industry, has largely been portrayed through hegemonic notions of normativity. Through simple distinctions between good and bad, a clear contrast is set up between a hero and an anti-hero. R. W. Connell suggests a hierarchy of masculinities, in the middle of which lies Marginalised Masculinity. This study constructs this marginalised masculinity as working-class. We contend that at the intersection of class and socioeconomic status lies a violent expression of masculinity accepted by the neo-liberal audience of Netflix. The Netflix original productions considered for this study exempt the working-class man from ethical obligations, in his conquest to liberate himself from his class position. Aspirational capitalism makes it okay to be ethically dubious, spawning a troubling masculinity. We use a three-step coding method to substantiate qualitative discourse analysis of Jamtara - Sabka Number Aayega (2020, Soumendra Padhi), Class (2023, Ashim Ahluwalia), and The White Tiger (2021, Ramin Bahrani).

Abstract

The swift advancement of vehicular edge computing (VEC) has made task offloading to roadside units (RSUs) an essential solution for addressing the computational demands of modern vehicular applications. However, efficiently partitioning and scheduling tasks across RSUs in dynamic environments remains a significant challenge. Traditional static or heuristic-based scheduling methods often struggle to adapt to dynamic vehicular networks, which leads to reduced resource utilization, increased latency, and reduced task completion rates, particularly in high-density scenarios. In this thesis, we explore the benefits of partitioned scheduling for computation offloading from multiple vehicles to RSUs, focusing on two key strategies: Fully Partitioned Scheduling (FP-Sched) and Hybrid Partitioned Scheduling (HP-Sched). These strategies are designed to handle the challenges posed by vehicle flow constraints, edge resource limitations, and bandwidth schedulability. By leveraging realworld vehicular datasets and simulations, we demonstrate that the proposed algorithms significantly improve task completion rates and reduce latency compared to traditional approaches. To further increase the adaptability and schedulability of vehicular tasks, we introduce HAPTA (Horizontal, Adaptive Learning-based, Partitioned, Timeslot-based Algorithm), a novel framework that leverages the Upper Confidence Bound (UCB) algorithm. HAPTA dynamically selects RSUs with optimal resource availability while considering vehicular mobility, task deadlines, and stringent latency constraints. The framework supports horizontal task partitioning, enabling the splitting of tasks into subtasks for efficient resource utilization across multiple RSUs. Experimental evaluations demonstrate that HAPTA outperforms existing heuristic-based methods, achieving up to 96.5% task completion in high-density scenarios while maintaining a substantial speedup over optimal scheduling frameworks. With the growing adoption of Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communication technologies, the proposed framework provides a robust solution for distributed computation offloading in next-generation intelligent transportation systems. This work contributes to the advancement of VEC by offering a scalable, adaptive, and efficient approach to task scheduling, paving the way for seamless integration of autonomous and connected vehicles in smart urban environments.

Abstract

As machine learning models are increasingly adopted in critical decision-making domains, ensuring their fairness has become a central concern—especially when these systems have the option to abstain from making a prediction. While abstain option classifiers reduce the risk of error by deferring uncertain decisions, they may inadvertently introduce new forms of unfairness, particularly when abstention rates differ across sensitive groups. This thesis investigates the fairness implications of abstain option classification and proposes methods to mitigate associated disparities. We present EQUISCALE, a unified framework for training fair cost-based abstain classifiers under group fairness constraints. Unlike prior work that focuses exclusively on coverage-based models and single fairness criteria, EQUISCALE introduces fairness-aware loss formulations that extend demographic parity (independence) and equalized odds (separation) to the abstention setting. The approach uses Lagrangian dual optimization to enforce fairness constraints during model training. EQUISCALE is the first method to incorporate multiple fairness criteria simultaneously in abstain classification and to apply the separation criterion in this context. We validate our approach through extensive experiments on five benchmark datasets, showing substantial reductions in fairness violations without compromising classification performance. This work advances the understanding of fair abstention and provides practical tools for deploying equitable machine learning systems in high-risk environments.