Dissertations - M Tech (CS)

Permanent URI for this collectionhttps://dspace.isical.ac.in/handle/10263/2147

These Dissertations were submitted in partial fulfilment of the requirements for the award of M TECH (Computer Science) Degree of Indian Statistical Institute

Browse

Now showing 1 - 20 of 617

2.5D Dual-Encoder U-Net for Lesion Segmentation in Chest CT Scans
(Indian Statistical Institute, Kolkata, 2025-06) Mukkara, Jagannath
Accurate segmentation of lesions in chest CT scans plays a vital role in diagnosing and monitoring pulmonary diseases such as COVID-19. In this, we introduce a novel 2.5D[1] dual-encoder U-Net model[2] that utilizes both the central slice and its neighboring slices to improve segmentation accuracy while keeping computational demands manageable. Our model incorporates residual connections[3] and feature fusion[4] to effectively merge multi-slice contextual information, overcoming the limitations found in traditional 2D and 3D methods. To ensure a reliable evaluation and avoid data leakage, we used patient-level data splitting. We validate our approach on a carefully curated chest CT dataset, showing enhanced segmentation performance and better generalization compared to standard U-Net models. Through extensive experiments, including ablation studies and visualizations, we demonstrate the advantages of combining 2.5D learning with a dual-encoder architecture for medical image segmentation tasks.
A Study of Prompt Tuning on Small Language Models(SLMs): A Controlled Benchmark and a Lightweight Instance-Aware Method
(2026-06-16) Sahith, Narkadamilli
Parameter-efficient fine-tuning (PEFT) adapts a frozen pre-trained language model by training only a small number of additional parameters. Among PEFT approaches, prompt tuning prepends trainable continuous vectors (soft prompts) to the input. A recurring finding in the literature is that prompt tuning is strongly scale dependent: it rivals full fine-tuning on very large models but lags on smaller ones. This dissertation studies prompt tuning specifically in the small-language-model (SLM) regime. We (i) re-implement a representative set of prompt-tuning methods—Prompt Tuning, P-Tuning v2, LoPT, DPT, DePT, ACCEPT, Residual Prompt Tuning, and PARA—within a single controlled harness, enabling a fair head-to-head comparison against full fine-tuning; (ii) propose IA-DePT, a lightweight instance-aware extension of Decomposed Prompt Tuning that conditions the short soft prompt on each input through a small, zero-initialised gate; and (iii) extend the benchmark beyond a single backbone and task, evaluating the full method suite on six backbone/task settings that span encoder–decoder (t5-small), encoder-only (BERT-base, RoBERTa-base, ELECTRA-small), and decoder-only (DistilGPT-2) architectures across the GLUE/SuperGLUE tasks RTE, WSC, CB, COPA, WiC, and MRPC. On RTE with t5-small, IA-DePT is the strongest parameter-efficient method in our benchmark (55.6% single-seed accuracy) and improves over its own base, DePT, by 6.5 points (53.6% vs. 47.1%, mean over three seeds) while adding only ≈16.9k parameters—a total trainable footprint of 0.05% of the backbone. Because the gate degrades exactly to DePT at initialisation, the comparison is a clean single-variable ablation. The cross-architecture study shows that the instance gate improves on DePT in five of the six settings on each setting’s primary metric (it ties or marginally regresses only on WiC, where every PEFT method sits at chance), so the benefit is broad but not universal. Our analysis characterises the accuracy/parameter trade-offs across method families, the strong effect of task difficulty on the small-model regime, and the role of instance-conditioning, including an honest discussion of why many prompt-tuning methods remain close to the chance baseline at this scale.
A Switch-Point-Aware Contrastive Approach to Sentiment Analysis of Hinglish Code-Mixed Text
(Indian Statistical Institute, 2026-06-15) Sahoo, Prasant Kumar
With the increasing use of social media in non-English-speaking regions, especially in India, people often use Romanized Hindi and English together in their online communication. In a single sentence, they frequently mix Romanized Hindi and English, creating code-mixed text. However, most multilingual transformer models are pre-trained primarily on monolingual data. As a result, NLP systems face challenges when processing code-mixed text, as a single word may be fragmented into meaningless subword pieces, making it difficult for the model to capture its semantic meaning accurately. In this dissertation, we propose a parameter efficient neural architecture consisting of three main components to address these challenges: First, there is a character-level CNN encoder, which handles spelling differences such as "nahi", "nahin", "nah", and "nai" through the chracter n-gram pattern. Next, there is a frozen XLM-R backbone(Conneau et al., 2019) , the top three layers, which are partly fine-tuned at a slower rate by which it provides rich cross lingual embeddings. Finally, there is a switch-point-aware bilingual gate that spots where the language label switches and blends two adapters using a learned gate weight.During training, it uses Supervised Contrastive Loss to learn better feature representations and Cross-Entropy Loss for classification. Since human annotators agreed on labels only 55% of the time, we use label smoothing to reflect this uncertainty and prevent the model from becoming overly confident in noisy labels. Evaluated on the SentiMix 2020 benchmark(Patwa et al., 2020), our proposed architecture achieves a weighted F1 score of 0.705, which outperforms the baseline model M-BERT (0.654 F1) and is comparable to fully fine-tuned transformer models while requiring only one-tenth of the trainable parameters.Adapter gate visualizations provide interpretable evidence that the gating mechanism captures linguistically meaningful codemixing structure. The architecture is designed to generalize to other code-mixed language pairs through its modular adapter design.
Access structures for an image database
(Indian Statistical Institute, Kolkata, 1992) Kuila, Sudhansu Sekhar
Acyclicity Tests in Classes of Dense Digraphs in Streaming Model
(Indian Statistical Institute, Kolkata, 2020-07) Kundu, Madhumita
Graph is a popular model to represent highly structured data which involves entities who have pairwise relations between them. In many applications, computing graph theoretic properties after modelling the entire dataset as graph, provides us interesting informations which gives us insights about the whole dataset. However, in case of application, the datasets in question can be so large that it's di cult to store in the main memory and the dataset can even be dynamic(can change with time). These days in so many applications, the algorithm that requires to solve the problem which takes massive dataset as input, has limitations on time as well as space taken to store the information. These constraints leads us for the development of new techniques. Streaming model of computation takes all these challenges into account and provides us solutions with limited resources in cost of accuracy. Graph stream is a sequence of imcoming edges and we are only allowed to insert(insertion only model) or both insert and delete(dynamic model) into an initially empty graph. Finally our objective is to nd out certain properties of the graph at the end of the stream which minimizes the amount of space the algorithm uses. Sometimes this algorithm needs to provide the trade of between the space usage and the time taken. There is a large volume work on undirected graphs in streaming model but the area of directed graph stream is a pretty unexplored. In this project, we study the problem of testing acyclicity in dense digraphs in semi-streaming model. Here the graph on n vertices is presented as a stream of edges and using O(n polylog(n))-space, we must determine if it is acyclic or not
Adaptive Spectral Trust Gate for Physics- Constrained Operator Learning
(2026-06-16) Chakraborty, Soham
Physics-informed machine learning improves the plausibility, data-efficiency and generalization of surrogate models by injecting prior physical knowledge into the learning process. The current approaches can be broadly divided into two main categories: soft constraints, which add a physics residual to the training loss but guarantee nothing at inference time, and hard constraints, which project the model output onto the constraint set exactly but apply the projection uniformly to every part of the signal — including parts that are dominated by noise, discretization error, or model mismatch, where the idealized physics is not actually trustworthy. This dissertation proposes the Adaptive Spectral Trust Gate (ASPINO), a mechanism that learns where to trust the physics. Operating in the Fourier domain on top of any surrogate model, a small gating network forms a per-mode convex combination of a data-driven soft path and a physics hard path. The gate is driven by features of the spectral coordinate and the spectral amplitude, so that it can apply the hard constraint in well-conditioned spectral regions and defer to the data-driven operator in regions corrupted by noise or aliasing. A single gate serves two very different hard paths - the linear Leray projection (incompressible flow) and a nonlinear rank-r SVD projection (massive-MIMO channel estimation). On the theoretical side, we give an empirical-Rademacher-complexity analysis: an unconditional safety floor — the gated class never exceeds the soft path it wraps — and, under a stated low-rank-transfer assumption, a capacity-reduction factor of 1 − ¯α (1 − √ρr), where ¯α is the fraction of capacity routed through the hard path. On Kolmogorov-flow denoising ASPINO is simultaneously the most accurate and near physical, dominating the unconstrained, hard and soft baselines; on ray-traced MIMO it improves a strong physics-informed baseline across all pilot budgets and SNRs without ever regressing. A third study, zero-shot super-resolution on the Poisson equation, confirms the discretization invariance of the gated construction. ASPINO is discretization-invariant and “plug-and-play” over the underlying operator.
Addressing class imbalance problems to improve animal detection through aerial image data
(Indian Statistical Institute, Kolkata, 2025-06) Koushal, Suryang
Monitoring animal populations in wildlife reserves is essential for conservation, especially for endangered species, but manual censuses are costly, risky, and logistically challenging due to vast, inaccessible terrains. Unmanned Aerial Vehicles (UAVs) with digital cameras provide a safer, scalable solution for collecting aerial imagery to estimate animal populations. However, semi-automated processing of these images faces significant challenges due to class imbalance in datasets, including foreground-background disparities, where background terrain dominates over sparse animal instances, and inter-class imbalances from uneven species representation and varied visual appearances (e.g., species, sizes, fur patterns) against diverse backgrounds like deserts or forests. These imbalances hinder Convolutional Neural Networks (CNNs) used for object detection, leading to inaccurate population estimates. This project addresses these issues using a dataset of 561 aerial images from Tsavo National Parks (March 2014) and Laikipia-Samburu Ecosystem (May 2015), collected by the Kenya Wildlife Service. We propose a clustering-based approach to categorize background terrain into distinct classes (e.g., desert, grassland), aiming to mitigate imbalances and improve animal detection accuracy in UAV imagery, supporting reliable, data-driven conservation strategies.
Administrative document processing
(Indian Statistical Institute, Kolkata, 2016) Chandra, Satish
Adversarial Attack on Neural Machine Translation System
(Indian Statistical Institute, Kolkata, 2019-06) Abijith, K P
Nowadays Deep Neural Network based solutions are deployed to solve numerous tasks. Thus, it has become absolutely important to study the robustness of these systems. Machine Translation is one of the popular applications of Deep Neural Networks. This thesis studies the robustness of Neural Machine Translation systems by generating adversarial examples with the objective to fool the model. Whenever there is a change in the source, i.e. when a word in the input sentence is replaced by an unrelated word, the translation system is supposed to re ect the changes while doing translation. These unwanted invariance learned by the model is undesirable. With intention to exploit this undesirable property learned by a Neural Machine Translation system we design an attack called: Invariance-based targeted attack. This attack introduces multiple changes(replacement of words) to the original input sentence, keeping the translation unchanged. In-order to facilitate the explanation of the design of the attack we introduce two methods: (i) Min-Grad method: To identify the position where a replacement of the word makes the least change in the translation, and (ii) Soft-Attn method: To search for a new word to replace, given a list of choices. The initial part of the report explain the preliminary explorations we did in-order to get some insights on how to do the problem formulation. These experiments are run on LSTM based models with single replacement policy. Using the learning from the rst part we extend the experiments to Transformer and BLSTM based models, which are considered as the state-of-the-art systems for machine translation.
Air-Writing Recognition
(Indian Statistical Institute, 2026-06-15) Shukla, Gaurang
Air-writing is the act of tracing characters or words in free space with a fingertip, recorded by a camera, giving a touch-free input modality for smart displays, augmented and virtual reality, and assistive interfaces. It is difficult because the finger never lifts: connecting strokes join adjacent letters with no pen-up signal to mark boundaries, and the same word varies widely in scale, position, and slant across writers. The WiTA benchmark of Kim et al. provides a large, person-disjoint dataset and a baseline that treats each clip as RGB video, recognised by a spatio-temporal 3D residual network trained with a CTC objective, reaching a character error rate (CER) of 0.292 on the English subset. The main goal of this dissertation was to improve on this error rate, which we achieve: we replace raw video with an explicit fingertip-trajectory sequence extracted from hand landmarks, fed to a Conformer encoder with a joint CTC/attention head. The resulting system attains a test CER of 0.219, improving on the published 0.292 of Kim et al. and 0.299 of Tan et al. by 15–27% relative.
Algorithm for mapping boolean network to LUT based FPGAs
(Indian Statistical Institute, Kolkata, 2001) Bhattacharyya, Jayasri
Algorithms and bounds in online learning
(Indian Statistical Institute, Kolkata, 2016) Sharma, Ankit
Algorithms for biological cell storing
(Indian Statistical Institute, Kolkata, 2010) Chatterjee, Soumyottam
Algorithms for Boundary Labeling of Horizontal Line Segments
(Indian Statistical Institute, Kolkata, 2019-06) Kurmi, Abhilash
In boundary labelling problem the target is to labeling a set P of n points in the plane with labels that are aligned to side of the bounding box of P . In this work, we investigate a variant of this problem. In our problem, we consider a set of sites inside a rectangle R and label are placed in the compliment of R and touches the left boundary of it. Labels are axis- parallel rectangles of same size and no two labels overlaps. We introduce a set V , called visibility , which is a set of subsets of labels correspond to points of sites. Before connecting site (say p) at point (say p1 2 p) with some label (say l), first we need to check weather subset of label correspond to p1 is in set V or not. If it is then we check the label l belongs to that subset of label or not. If it contains that label then we can join site to the label, otherwise not. In our problem we used po-leaders, that is starting from site it is parallel to the side of R where its label resides and then orthogonal to that side of R. We considered various geometric objects as sites, such as point, same length horizontal segment, different length horizontal segments. As a solution, we derive a dynamic algorithm that minimizes the arbitrary cost function and give us planar solution where sites connects to labels by po- leaders and induces a matching such that no two po-leader intersects, also no two leaders shares common site (or label) and every leader satisfies visibility V . For points as sites, our dynamic algorithm runs in O(n3) time and optimizes the cost function. This running time also same for the case of unit length horizontal line segments as sites. Then we taken arbitrary length horizontal segment, algorithms runs in O(n4) time. We assumed that only one end point of any horizontal line segment can be used to connect label (by po-leader).
Algorithms for finding isomorphic subgraphs
(Indian Statistical Institute, Kolkata, 2001) Giri, Pradeep Kumar
Algorithms for thinning of gray level bengali script
(Indian Statistical Institute, Kolkata, 1992) Sarkar, Manish
Algorithms on geometric graphs
(Indian Statistical Institute, Kolkata, 2010) De, Minati
American Sign Language Recognition and Analysis Using Deep Learning
(2026-06-19) Soni, Saurabh Kumar
In this work I build a system that recognizes isolated American Sign Language (ASL) words, and I use it to ask one fairly direct question: when training data is scarce, is it better to look at the video pixels or at the geometry of the signer’s body? To find out, I train two very different models on exactly the same clips. The first is appearance-based. Every frame is run through standard preprocessing and a ResNet50 backbone pre-trained on ImageNet, which turns it into a 2048-dimensional feature vector, and a Bidirectional LSTM then reads that sequence over time. The second model never sees a pixel. It works only on Media Pipe key points, the tracked coordinates of the body and the two hands, and feeds them to a Transformer encoder. So both models have to learn the same two things, the shape of the hands in each frame and the way those shapes move across frames, and both are trained, validated and tested under one identical protocol. What I care about throughout is a recognizer that is accurate but still light enough to be useful in practice, so it could eventually make communication a little easier between people who sign and people who do not.
An Efficient Hierarchical Deployment Of Sensors For K-coverage In Planner Wireless Sensor Network
(Indian Statistical Institute, 2026-06-17) Singh, Abhay Raj
Ensuring reliable sensing coverage is a fundamental challenge in wireless sensor networks (WSNs), particularly when multiple sensors are monitoring each location to provide robustness against node failures. In this work, we address the problem of deterministic k-coverage in planar WSN by proposing a hierarchical triangular lattice-based deployment strategy that organizes sensor locations across di↵erent refinement levels and guarantees coverage of every point in the sensing domain by at least k sensors. Each lattice is three-colorable, and selective activation of color classes ensures adjustable coverage guarantees. We prove that activating a single color class at refinement level t guarantees at least 4t-coverage inside a triangular region. Using a base-4 decomposition of the required coverage level k, we construct a deployment scheme that achieves arbitrary k-coverage while minimizing the number of sensors. For finite irregular hexagonal (IRH) domain, we drive closed-form expressions for the exact number of sensors and minimum sensor spacing required to ensure k-coverage. Analytical comparison with existing IRH edge-overlap, IRH inner-diamond and square-band based deployment strategies show that the proposed method either reduces the required number of sensors or increases the minimum sensor spacing while maintaining the similar coverage guarantees. The analytical results are also validated against extensive simulations. The proposed framework provides a constructive and scalable approach for efficient sensor deployment in large-scale wireless sensing systems.
An Empirical Study of RLVR Fine-Tuning for Mathematical Problem Solving in LLMs
(Indian Statistical Institute, 2026-06-19) Konnur, Rashmi
Large language models have shown immense improvement in coding and math performances thanks to reinforcement learning boosted algorithms. However, its true impact on broadening the reasoning and analytical capacities of an LLM is still contended. In this dissertation, we outline the foundations of Large Language Models, and delve into Reinforcement Learning with Verifiable Rewards (RLVR). We discuss various strategies to efficiently manipulate memory during a fine tuning update. We finally perform RLVR fine-tuning techniques on different models with varied use cases and compare their performances, which corroborate the efficiency of RLVR.