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
62 results
Search Results
Item Dynamic Property Ordering for Efficient Multi-Property Bounded Model Checking(Indian Statistical Institute, 2026-06-16) Kumar, VivekFormal verification plays a critical role in ensuring the correctness of modern hardware designs. As the complexity of digital systems increases, designs are often associated with a large number of verification properties that must be analyzed within limited computational resources. In conventional multi-property bounded model checking (BMC), all properties are verified simultaneously. While this approach enables parallel analysis, difficult properties can consume a disproportionate amount of resources, causing simpler properties to be delayed and reducing the overall efficiency of bug detection. This thesis presents dynamic property ordering techniques for efficient multi-property verification using SAT-based bounded model checking in the ABC verification framework. The central idea is to verify properties individually and dynamically prioritize them based on their observed verification progress, allowing computational resources to be directed toward properties that are more likely to yield results within a given time budget.Two dynamic property ordering algorithms are proposed. The first algorithm, ALG1, employs a round-robin style strategy in which unsolved properties are periodically reordered according to the maximum verification depth (frame) reached, prioritizing properties that demonstrate greater progress. The second algorithm, ALG2, adopts a priority-based scheduling approach where each property’s priority is determined by its verification rate, measured as frames explored per second. Properties with higher progress rates are allocated greater verification resources. The proposed approaches are evaluated on benchmark suites from the Hardware Model Checking Competition (HWMCC) 2012 and 2013 and compared against two baselines: the conventional ABC multi-property verification method and an Equal Time Bounding (ETB) strategy that distributes the available verification time equally among all properties. Experimental results demonstrate that dynamic property ordering significantly improves verification efficiency. Both ALG1 and ALG2 solve more properties and achieve greater verification depth within the same time budget, while also accelerating bug discovery. Across the benchmark set, the proposed methods provide improvements exceeding 40% over the baseline approaches in key performance metrics. The results demonstrate that dynamic property scheduling is an effective technique for improving the scalability and effectiveness of multi-property bounded model checking, offering a practical solution for faster bug detection and enhanced utilization of verification resources.Item Deep Reinforcement Learning with Directed Asymmetry and Kolmogorov-Arnold Networks for Dismantling Interdependent Multiplex Networks(Indian Statistical Institute, 2026-06-16) Dev, SoumyajitIdentifying the minimum-cost node-removal sequence that fragments a complex network - the network dismantling problem is NP-hard and central to infrastructure resilience. In interdependent multiplex networks, this difficulty is compounded by cascading cross-layer failures. While deep reinforcement learning (DRL) agents utilizing graph neural network (GNN) encoders achieve near-optimal dismantling, current state-of-the-art architectures suffer from two critical limitations. Topologically, existing agents strictly assume undirected edges, rendering them inapplicable to directed systems - such as supply chains or gene regulatory cascades - where failure propagation is fundamentally asymmetric. To resolve this, we propose Disassembling Directed Interdependent Networks (DDIN). DDIN introduces an asymmetric GraphSAGE encoder that explicitly separates incoming influence from outgoing control aggregations, paired with a multi-relational attention mechanism for cross-layer dependency fusion. Evaluated zero-shot on five real-world directed networks, DDIN achieves a 16-23% reduction in the Area Under the Dismantling Curve (AUDC) over heuristic baselines. Functionally, existing GNN encoders parameterise message-passing through multi-layer perceptrons (MLPs). These fixed-affine projections lack the capacity to model the non-smooth, high-frequency vulnerability patterns governing cascading fragility in scale-free topologies. We address this by proposing Kolmogorov-Arnold Reinforcement Learning (KARL), the first architecture to embed learnable univariate functions into an off-policy DRL combinatorial optimiser. KARL features a residual B-spline KAN encoder to stabilise gradient norms during deep message-passing, a fully KAN-parameterised cross-layer attention module (KANformer), and an orthogonal Chebyshev-KAN action-value head to mitigate boundary oscillations under non-stationary temporal-difference targets. Evaluated zero-shot on six real-world undirected multiplex networks, KARL yields a 21.96% average AUDC improvement over state-of-the-art baselines, alongside emergent structural interpretability. By independently resolving the directed topology limitation and the affine expressivity bottleneck, this dissertation establishes a robust architectural foundation for structurally faithful network dismantling models.Item Adaptive Spectral Trust Gate for Physics- Constrained Operator Learning(2026-06-16) Chakraborty, SohamPhysics-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.Item Predictive importance sampling based coverage verification for multi uav trajectory planning(Indian Statistical Institute, 2026-06-23) Ghosh, SnehashishIn next-generation wireless networks, unmanned aerial vehicle (UAV) networks are emerging as a promising solution for ultra-reliable low-latency communication (URLLC). A key challenge in millimeter-wave UAV networks is ensuring that mobile users are always in line-of-sight (LoS) coverage, since the current snapshot-based trajectory planning approach does not consider the mobility of the users during the decision interval, resulting in disastrous LoS gaps. For continuous coverage verification, standard uniform sampling is too computationally expensive, as it would need a large number of samples to estimate rare failure events that have latencies that are not suitable for real-time requirements. In this work, we introduce a Predictive Importance Sampling (PIS) framework that significantly decreases sample complexity by focusing verification efforts on regions where failure is predicted. Specifically, we propose a Long Short-Term Memory Mixture Density Network (LSTM-MDN) architecture to learn multimodal user trajectory distributions and introduce a defense approach based on mixture sampling to handle the robustness against the prediction error. We show that PIS yields unbiased failure probability estimates that have lower variance than uniform sampling. We then combine PIS with Multi-Agent Deep Deterministic Policy Gradient (MADDPG) to perform coordinated multi-UAV trajectory planning based on an energy-aware multi-objective reward function that balances throughput, coverage, fairness and energy consumption. Based on the simulation results, our proposed method improves the coverage rate, throughput and verification latency in comparison with three state-of-the-art methods, thus enabling proactive coverage management for URLLC-aware UAV networks.Item Telluric Correction of M-dwarf Stars using Machine Learning(Indian Statistical Institute, 2026-06-17) Rana, SayakThe study of M-dwarf stars is of prime scientific interest to us because of their closer habitable zones and the favorable conditions they offer for exoplanet detection. However, telluric contamination of the ground-based spectra results in sharp absorption lines, which makes their study cumbersome. Removing this contamination is necessary for estimating key stellar parameters. The central contribution is a one-dimensional Convolutional Neural Network (CNN) that retrieves the four atmospheric parameters governing telluric absorption: pressure, temperature, humidity, and airmass. These predicted parameters are passed to Telfit which produces an estimated telluric spectrum. The observed spectrum is then divided by this estimated telluric spectrum to obtain the telluric corrected spectrum. As this network is trained exclusively on synthetic spectra, a domain gap exists at inference. Two domain-adaptation strategies are evaluated: a CycleGAN following the Cycle-StarNet framework for explicit synthetic-to-real translation and a Domain-Adversarial Autoencoder (DAAE) that learns domain-invariant spectral representations. The Domain-Adversarial Autoencoder (DAAE) achieves the lowest loss against a telluric corrected CARMENES reference spectrum outperforming all other model variants. The CycleGAN fails due to discriminator collapse under severe class imbalance between the number of real and synthetic spectra.Item Daily Rainfall Forecasting over West Bengal with Spatio-Temporal Graph Networks(Indian Statistical Institute, 2026-06) Ghosh, RisheekDaily rainfall is hard to forecast where most days are dry, a few days are very wet, and almost all of the rain arrives in one season. This dissertation studies day-ahead rainfall forecasting over West Bengal, India, using models that learn from both the temporal and the spatial layout of measuring stations. We first build up a forecasting model in stages, from simple models that look at one station’s past to a graph-based model that links nearby stations, and we identify the strongest deterministic forecaster among them. We then ask whether the choice of input data changes the story, by comparing rain-gauge observations against a reanalysis product on the same stations; the reanalysis turns out to be a smoothed stand-in that misses the sharpest days. Looking closely at the best model, we find it behaves like an estimator of the average rainfall for each day, which is why it cannot place the rare heavy events: almost all of the error comes from a small number of very wet days, and a squared-error objective drives the model to predict too little on exactly those days. We show that self-supervised pretraining and related representation tricks do not move this ceiling. Finally, we change what the model predicts: instead of a single number we predict a full range of possible rainfall values. This probabilistic model is well calibrated and produces useful, reliable warnings for heavy-rain thresholds that the single-number model could never flag, at the modest cost of a slightly worse typical-day error. The contribution is a clear, honest account of where standard spatio-temporal deep learning succeeds and fails for daily rainfall, and a simple change of formulation that recovers useful accuracy on the events that matter most.Item An Empirical Study of RLVR Fine-Tuning for Mathematical Problem Solving in LLMs(Indian Statistical Institute, 2026-06-19) Konnur, RashmiLarge 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.Item Efficient Performance Recovery of Pruned LLMs for End Devices(Indian Statistical Institute, 2026-06-23) Paul Vardhan BethapudiLarge Language Models have shown strong performance across reasoning, language understanding, and generation tasks, but their computational and memory requirements make deployment on low resource devices difficult. This dissertation studies efficient performance recovery of pruned LLMs, focusing on whether a pruned model can regain useful task performance through parameter-efficient fine-tuning while preserving the benefits of compression. The work uses Llama-3.2-3B-Instruct as the dense reference model and applies activationaware Wanda pruning at multiple sparsity levels(20%, 30% and 50%). The pruned models are evaluated on language modeling and reasoning tasks using WikiText-2 perplexity and GSM8K (Grade School Math 8000 problems) accuracy. Adaptation is performed using efficient parameter adaptation methods like QLoRA and DoRA. Additionally, the thesis evaluates the balance between sparsity, precision, perplexity, speed, memory usage, and device compatibility. Based on the results obtained from the experiments conducted, we can safely say that the adapter recovery technique improves the performance of sparse models. DoRA performs better in GSM8K recovery compared to QLoRA recovery at sparsity of 20%, 30%, and 50%. Additionally, the WikiText-2 perplexity for QLoRA is slightly lower than DoRA at 30% and 50% sparsity. The largest recovery gain occurs at 50% sparsity, where GSM8K accuracy improves from 0.1800 to 0.4508 using DoRA recovery. However, the best final recovered GSM8K accuracy is obtained at 20% sparsity with DoRA.Item Leveraging Spatial Statistics for Domain Adaptation of Vision Language Models in Medical VQA(Indian Statistical Institute, 2026-06-23) Raj, HimanshuRecent advances in Vision–Language Models (VLMs) have demonstrated strong performance in Medical Visual Question Answering (Medical VQA) task. Although they perform very well within their domains, these models often experience issues with their generalization ability on unknown clinical distribution data because of different imaging technologies and patient groups used in various medical facilities. Generalization problems faced by these models make their practical application in the field of VLM-based medical VQA systems rather difficult. To overcome this limitation we proposed our method named Spatial Semantics Aware Domain Adaptation (SSADA), which is an integrated framework that combines both finetuning and prompt-based in-context learning for domain adaptation. Our proposed approach, SSADA, includes the following three important components: (i) Mask-Aware Finetuning (MAFt) to make localization aware finetuning, (ii) Anatomy Aware Instance Normalization (AAIN) for handling intensity or distribition shift, and (iii)Weighted Multi-Modal Example Retrieval (WMMER) for semantically consistent example selection during inference. We evaluate the proposed framework on three publicly available Medical VQA benchmarks, SLAKE, VQA-Med 2019, and OmniMedVQA–RadImageNet, under cross-domain settings and compare it against standard finetuning techniques. Experimental results demonstrate the effectiveness of SSADA in improving cross-domain generalization.Item Air-Writing Recognition(Indian Statistical Institute, 2026-06-15) Shukla, GaurangAir-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.