Abstract

Modern Optical character recognition (OCR) systems increasingly rely on deep visual sequence
models, yet they are often evaluated mainly through final recognition accuracy. This leaves an important
question underexplored: how do such models learn, stabilize, and exhibit persistent errors over the
course of training? This question is especially relevant for Indic scripts, where visual text recognition is
shaped by complex orthographic structures such as aksharas, modifiers, conjunct-like forms, and dense
character compositions.
This thesis studies the learning behavior of two representative OCR models, CRNN and PARSeq,
using a five-language corpus covering Hindi, Bengali, Tamil, Telugu, and English. The main chapters
present representative analyses, and the same qualitative trends were observed in the remaining languages. The experiments are conducted on cropped printed word images, allowing the analysis to focus
on recognition behavior at the word level. Rather than comparing models only by final CER and WER,
the thesis examines their behavior across four complementary dimensions: training data scale, sequence
length, subword structure, and optimization geometry.
The results show that learning behavior varies meaningfully across model family, data regime, sequence length, and script structure. Under the tested settings, CRNN is more sensitive to reductions in
training data, while PARSeq remains comparatively more stable. Short words are learned faster than
long words, showing that sequence length remains an important source of recognition difficulty. Frequent and structurally simpler subword patterns stabilize earlier than rarer and more compositionally
involved patterns. Loss landscape visualizations further suggest that training moves from sharper and
more irregular regions toward smoother and more stable basins, with fuller data settings reaching such
regions more readily.
Together, these findings show that final recognition scores alone do not fully characterize OCR model
behavior. The analysis can support more informed model selection, especially when data availability,
script complexity, and word length vary across recognition settings. The findings may also inform
dataset design by indicating where additional examples are likely to be useful. By focusing on learning
trajectories, this thesis offers a structured way to evaluate OCR models beyond final accuracy, with
attention to how their behavior changes with data availability, word length, script structure, and training
stability.

Abstract

Small centrifugal pumps under 10 HP form the operational backbone of community-scale water supply across rural and peri-urban regions of the developing world, yet they are monitored exclusively
through periodic manual inspection that cannot detect developing faults before they cause service disruption. Digital Twin technology offers a well-established solution, but all existing implementations
target large industrial assets with infrastructure requirements that are economically and technically incompatible with community-scale installations. This thesis bridges that gap by designing, deploying,
and quantitatively validating the first open-source, physics-grounded operational Digital Twin for a
3.7 kW three-phase induction motor driving a live campus water pump, built entirely from commodity
sensors, open-source middleware, and standard cloud infrastructure.
The system integrates a Simulink-based virtual motor model covering coupled electrical, mechanical,
and thermal dynamics with Eclipse Ditto as a Digital Twin (DT) middleware, MongoDB for long-term
telemetry archival, and ThingSpeak for data storage and later integration with MATLAB, all interconnected via application-layer protocols operating over TCP/IP protocols over campus Wi-Fi. Temperature, current, and three-axis vibration data streamed from an ESP32 sensor gateway at five-second
intervals are continuously compared against model predictions; when the weighted composite RMSE
exceeds a statistically derived threshold for three consecutive cycles, the system issues a fault alert and,
under confirmed critical conditions, autonomously shuts down the motor through a solid-state relay,
completing a partially closed-loop operational Digital Twin without any proprietary software or dedicated industrial network.
A four-month deployment from August to November 2024 accumulated 160,000 sensor datapoints
and validated the system against four quantitative design objectives, all of which were met. The
Simulink model achieved normalised prediction errors of 3.2% for motor body temperature and 4.8% for
phase current, with end-to-end latency of approximately 8 seconds satisfying the SCADA supervisory
monitoring standard of below 10 seconds. A Random Forest classifier trained on heuristically labelled
fault data, requiring no documented fault history, achieved precision 0.84, recall 0.98, F1 score 0.90,
and AUC 0.79 across four fault classes; a comparative regression analysis of eight ML-based algorithms
further confirmed that the Random Forest Regressor best characterises the motor's thermal-electrical relationships (R2 = 0.77, MAE = 0.70 ◦C). An 11% reduction in normalised energy consumption was
observed over a 30-day pre/post-deployment comparison (p = 0.032), and a structured usability trial
with five non-technical facilities staff yielded a mean satisfaction score of 4.3 out of 5, with all operational tasks completed within a 15-minute training limit.

Abstract

Large language models now produce fluent text across domains, creating a need for systems
that can distinguish human-authored text from machine-generated text and, when needed, identify the generator. We study whether RoBERTa's multi-layer representations can support both
goals under distribution shift. In our initial SemEval study, we proposed weighted layer averaging (WLA), which learns scalar weights over RoBERTa layer representations, and AdaLoRA,
which adapts the encoder through parameter-efficient low-rank updates. The WLA component
follows the view that different encoder depths can expose different contextual features.
We extend that inquiry from binary detection to generator attribution. The main finding is
task-conditional. For binary detection, aggregation can support a broad human-versus-machine
boundary. For generator attribution, the same kind of compression appears less suitable. Further experiments indicate two practical risks: WLA layer averaging may blur layer-specific
generator cues, and the tested low-rank adaptation setup may be too restrictive for six-way
source identification under domain shift.
To address this compression problem, we evaluate layer-slice concatenation, which preserves
selected RoBERTa layers as independent channels before classification. The strongest reported
region consists of slices ending at layer 9: layers 5-9 and 6-9 reach 0.9323 test accuracy on
OUTFOX, with layers 2-9 marginally lower at 0.9319 and effectively tied after rounding. These
results exceed the plain RoBERTa attribution anchor of 0.8730 accuracy and suggest that
generator attribution benefits from preserving layer channels rather than using early scalar
compression. Overall, robust detection and generator attribution require different architectures:
aggregation for broad invariance, preservation for fine-grained source identification.

Abstract

This thesis investigates the dynamics of answer formation in Large Language Models (LLMs) along
two dimensions: vertically across network layers, and horizontally across sequential reasoning tokens.
To evaluate vertical evolution, we apply layer-wise linear probes to open-weight models on a controlled character-counting task. The probing results show that early and middle layers reliably encode
sub-token character counts. Late-stage network components, specifically the penultimate and final multilayer perceptrons, actively suppress this information and default to degenerate prediction strategies. The
models thus fail at utilization, despite having the required representational capacity.
For horizontal evolution, we introduce Forced Answer Completion (FAC) to extract a model's intermediate answer commitment at each step of its chain-of-thought trace. Across multiple-choice, numeric,
and search tasks, FAC shows that models decide their final answer long before the reasoning trace concludes. After filtering out Transient Answer Flips (TAFs) with a causal smoothing operator, we find
that nearly half of all reasoning tokens are generated after the model's decision has stabilized. These
late-stage tokens function primarily as post-hoc justification.
These findings enable direct interventions in LLM inference. Vertical tracking isolates suppressed
knowledge, providing targets for structural adjustments. Horizontal tracking supports lightweight earlystopping gates by training linear probes on intermediate hidden states, we predict when an answer has
stabilized and halt generation. Testing on the Qwen3 architecture demonstrates that this early-stopping
method saves hundreds of tokens per query with a marginal drop in accuracy

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing, yet their application to Machine Translation (MT), particularly for lowresource languages, remains an active area of research. This thesis investigates the adaptation
of LLMs for machine translation involving English and Indian languages, addressing challenges
in assessment, fine-tuning, alignment, evaluation, and inference efficiency.
We begin by assessing the translation capabilities of popular LLMs (LLaMA, Mistral, Bloom)
for 22 Indian languages under zero-shot and in-context learning settings. Our findings reveal
that while raw LLMs exhibit limited translation quality, parameter-efficient fine-tuning using
LoRA significantly improves performance, with two-stage fine-tuning achieving results competitive with specialized MT systems.
Building on these insights, we investigate fine-tuning strategies for extremely low-resource
Indic languages through participation in the WMT 2024 shared task. Our experiments across
Assamese, Mizo, Khasi, and Manipuri demonstrate that fine-tuned LLMs substantially outperform models trained from scratch, and that multilingual approaches consistently surpass
monolingual ones.
We then explore data-efficient alignment strategies using preference optimization. Our Progressive Perturbation approach with Kahneman-Tversky Optimization (KTO) demonstrates
that targeted perturbations of training data can enhance translation quality without requiring
extensive parallel corpora.
To address the challenge of evaluating MT for low-resource languages, we conduct a comprehensive analysis of statistical and neural evaluation metrics, examining their stability, consistency, and correlation with human judgments.
Finally, we tackle inference efficiency through DistillBeam, a multi-trajectory knowledge
distillation framework for speculative decoding. By aggregating supervision from multiple
high-probability teacher trajectories, DistillBeam achieves wall-clock speedups of 35-65% over
autoregressive decoding across 20 languages.
This thesis contributes practical methodologies and empirical insights for deploying LLMbased translation systems for Indian languages, balancing translation quality with computational efficiency.

Abstract

Deployment of Face Recognition (FR) systems on edge devices has seen significant growth, driven
by advances in hardware and neural architectures. However, tailoring State-of-the-art (SOTA) FR solutions to specific edge environments remains difficult and largely unexplored. Although benchmark data
exists for some combinations of model, hardware, and frameworks, it does not scale to the evaluation of
custom or modified architectures. Furthermore, theoretical metrics such as Floating-Point Operations
(FLOPs) are often poor predictors of real-world performance, as edge inference throughput is frequently
constrained by system-level factors including data movement, memory access, and pipeline overheads.
This limitation arises from a combinatorial design space spanning architectural variants, input resolutions, deployment frameworks, and hardware configurations, making exhaustive benchmarking infeasible. An analysis of recent FR literature performed in this thesis reveals that while substantial
innovation is concentrated in loss functions, training strategies, and data augmentation, the underlying Convolutional Neural Network (CNN) backbone is routinely treated as a fixed abstraction. These
backbones are typically inherited from large-scale image classification, where models are optimized to
separate visually distinct categories, whereas FR requires learning discriminative embeddings that distinguish between highly similar identities. Consequently, the exploration of lightweight, task-specific
architectures for edge deployment remains limited.
To address this gap, this thesis models the relationship between network architecture and inference
throughput in an edge deployment setting, demonstrating that throughput follows a predictable structure
when grouped by architectural families. Building on this modeling, this work proposes a data-efficient,
architecture-aware framework to estimate the throughput of custom FR models from a sparse set of
empirical measurements, with predictions that are explainable and exhibit errors small enough to be
treated as standard on-device observational noise.
Further, Face Anti-Spoofing (FAS) pipelines are evaluated to quantify the trade-offs between inference throughput and error rates, highlighting deployment challenges arising from specialized operations that do not map efficiently to standard edge-acceleration frameworks. In parallel, empirical experimentation on FR backbones is conducted to characterize the trade-off between discriminative power and
inference throughput, and its impact on deployability on edge hardware, revealing non-linear scaling
behavior and throughput saturation driven by data pipeline constraints.
Collectively, these findings address a key gap in FR deployment: the absence of a scalable mapping
between architectural design and on-device inference performance for custom models. By establishing
an explainable mapping from model architecture to throughput, the proposed framework enables systematic performance estimation without exhaustive benchmarking, supporting hardware-aware design
and exploration of custom architecture for FR. To the best of the author's knowledge, this is the first
work to directly address this problem. The complete benchmarking suite and source code are publicly
available at https://github.com/AyanBiswas19/Resource_Efficient_FR.

Abstract

The ability to comprehend and trace program execution is a foundational skill in computer science.
Yet, it remains a significant hurdle for novices. Existing pedagogical methods often prioritise tracking
variable states, leaving a critical gap in teaching the explicit control-flow path of a program's execution.
This thesis introduces and evaluates a novel pedagogical framework designed to address this gap by
making control flow the central object of study. The framework consists of a simplified subset of Python
and a set of six scaffolded, hand-drawn artefacts that guide students in representing a program's static
structure and dynamic execution.
The framework was implemented as a large-scale classroom activity in a first-year programming
course with 368 students. Detailed analysis of the artefacts constructed by 84 students yielded a mean
score of 74.56%, demonstrating the framework's efficacy as a structured methodology for tracing. Exploratory factor analysis of the scores identified four latent variables that influence student performance:
Structural Abstraction Proficiency, Sequential Control Flow, Condition-Dependent Control Flow, and
Procedural Control Flow. Furthermore, the study identified prevalent misconceptions, particularly related to procedural control transfer and error handling. These findings, along with the observed limitations, provide insights into the mental models students develop about program execution and offer a
foundation for refining the teaching of control flow in program execution.

Abstract

The Internet of Things (IoT) has revolutionized various fields by enabling the collection of data
in real-time and interacting remotely with physical systems. In the education sector, one of the most
impactful uses of IoT is remote labs, which enable students to perform experiments virtually. This
has been particularly impactful during the COVID-19 pandemic or for those who are located remotely.
Though remote labs are becoming extremely popular for physics and electronics due to their relatively
simpler requirements, chemistry has not yet found a suitable place due to the complexity involved. In
addition, existing remote labs are not focused on creating a deep level of reflection during learning.
This thesis presents a two-pronged approach to bridging the gap between traditional and remote
laboratory learning by focusing on chemistry education and pedagogical enhancement. First, it details
the design and implementation of a fully retrofitted, IoT-enabled titration system. The proposed system
has been implemented on the RLabs platform at IIIT Hyderabad and provides a facility for users to
conduct acid-base titration experiments virtually using a web-based platform. It consists of hardware
components that help in precise measurement of fluids, visual feedback mechanisms, and automatic
cleaning facilities. This helps users conduct repeated experiments without any physical interference.
The experimental setup has been tested for accuracy, usability, and cost-effectiveness, demonstrating its
potential as a scalable solution for remote chemistry education.
The second part of this work addresses the pedagogical shortcomings of existing remote labs, which
are mostly limited in their instruction and lack active engagement with learners during the experiment.
To fill this gap, a large language model (LLM)-powered assistant is integrated into the RLabs framework. Grounded in a dual-theoretical foundation of Kolb's Experiential Learning Cycle and Bloom's
Taxonomy, the assistant facilitates active experimentation, conceptual reflection, and assessment across
multiple cognitive levels. It includes context-based tutoring, structured viva examination, and experiment completion verification, mimicking the role of a human lab instructor. The system was evaluated
through user studies that benchmarked candidate language models and quantified the impact of staged
LLM-assisted guidance, showing clear improvements in learner performance across all six levels of
Bloom's Taxonomy.
These contributions show a marked advancement in the development of remote laboratory systems
by bringing together robust experimentation via IoT and intelligent tutoring. The contributions can be
seen as a step towards the development of more holistic and capable remote laboratory systems that can
effectively mimic in-person learning while being accessible to a wide audience.

Abstract

Speech prominence refers to the relative emphasis placed on certain units in speech to express the
communicative intent. This prominence is realized at two different levels. One, at the syllable level,
where one syllable within a word is made more prominent than the rest, referred to as lexical or syllablelevel prominence. Lexical prominence is generally produced correctly by native speakers and is commonly incorporated into speech synthesis through pronunciation lexicons, but non-native speakers frequently misplace it due to the influence of their native language, motivating the need for automatic
lexical stress detection. Two, at the word level, where certain words within an utterance are emphasized more than others, commonly referred to as sentence or word-level prominence. Unlike lexical
prominence, automatic detection and natural generation of sentence prominence remain challenging
problems for both native and non-native speech. Despite their importance, existing prominence modeling approaches continue to face several limitations. First, most methods rely on handcrafted prosodic
features, such as pitch, energy, and duration, which may not fully capture prominence-related information. Second, prominence modeling at both syllable and word levels typically depends on accurate
boundary annotations, obtained either through costly manual labeling or error-prone forced alignment.
Third, training these systems requires expensive prominence labels, limiting scalability to low resource
language backgrounds.
In this thesis, these limitations are addressed progressively, moving from supervised to unsupervised learning and from boundary-dependent to boundary-independent modeling across both linguistic
levels. At the syllable level, it is first demonstrated that self-supervised speech representations provide a substantially stronger foundation for prominence modeling than handcrafted features, with sequential modeling further improving performance by capturing inter-syllable dependencies. The dependence on boundary annotations is then removed through hierarchical temporal compression, introducing
boundary-independent frameworks that match and often exceed boundary-dependent counterparts. The
reliance on prominence labels is further addressed through Post-Net and Post-Net2.0, which leverage
pseudo-label generation and linguistic constraints to enable effective unsupervised prominence detection. A fully unsupervised and boundary-independent framework, SRAMA, is proposed, combining
Adaptive Local Monotonic Attention (AMA) with iterative deep clustering, approaching supervised
performance without requiring boundaries, transcriptions, or prominence labels. Finally, the scope is
extended to the word level through ProTTS, which integrates prominence discovery directly into SOTA
TTS, namely, FastSpeech 2, using word-level prominence inferred from prosodic predictions to guide
synthesis -- producing more natural and expressive speech and bridging the two central applications of
this thesis: automatic speech prominence detection and expressive speech synthesis.

Abstract

Despite negation being one of the core elements in any language, it remains a challenging
phenomenon for modern Large Language Models (LLMs). Recently, there have been growing efforts to evaluate how models handle negation. However, the existing probing datasets
are mostly English-centric, leaving morphologically rich languages largely unevaluated. To facilitate evaluation for Indian languages, especially Telugu, which expresses negation not only
through separate words but also through morphemes attached directly to words, we present
the NEGTEL benchmark. As part of this work, we first compile and verify an inventory
of Telugu negative morphemes and introduce NEGSEED, a manually curated set of example
sentences covering the full range of morphological negation patterns in Telugu. Building on
this resource, the NEGTEL benchmark is a test suite that contains five tasks: Negation Detection, Translation, Paraphrase Detection, Sentiment Classification, and Polarity Flipping. These
tasks are designed to evaluate models at increasing levels of linguistic and logical complexity,
starting from whether models can detect the presence of negation, to whether they can preserve
it during translation, distinguish negated sentence pairs, correctly classify sentiment in the presence of negation, and actively manipulate negation by flipping the polarity of a sentence. The
test suite is designed grounded in detailed linguistic analysis and includes annotations of different negation types. We evaluate a diverse set of multilingual LLMs on this benchmark, and
our evaluation reveals that most models struggle significantly with Telugu negation across all
tasks. A recurring finding is that standard automated metrics remain high even when models
fail semantically, highlighting the need for human evaluation when assessing negation-related
capabilities.