Computer Engineering Commons^™

Physiotrack: A Gamified Physiotherapy System, Pranavi Chaturvedula

Master's Projects

Traditional physiotherapy methods tend to be non-interactive and provide little to no personalized instruction, even though physiotherapy is critical to stroke recovery. This thesis explores a fully adaptive, sensor-based, feedback architecture intended for stroke patients which remotely supervises movement and personalizes exercises enabled by multimodal sensors. The system uses filtering and windowed segmentation of accelerometer and skeletal data to compute features like jerk, speed, and joint movement angular range. A game engine applies accelerometer and skeletal features together with optimized, lightweight ML models to drive adaptive feedback, scoring, and difficulty adjustment. The architecture supports responsive continuous sensor streaming within the …

Galora: A Lightweight Graph-Aware Llm Framework For Node Classification On Text-Attributed Graphs, Mayur Choudhary

Master's Projects

With the exponential rise of language models (LMs) and their potential to understand semantic relationships, large LMs are being used across a wide range of applications. Text-attributed graphs (TAGs) are one notable example where LLMs can be combined with Graph Neural Networks (GNNs) to enhance node classification results. TAGs associate textual content with each node and are commonly seen in various domains such as social networks, citation graphs, recommendation systems, etc. Effectively modeling TAGs would enable deeper insights into different aspects of the graph and improve decision-making in relevant domains. We present GaLoRA, a parameter-efficient framework to integrate structural information …

Extraction Of A Knowledge Graph Of Biomedical Relationships, Brian Tran

Master's Projects

Rapid release in biomedical literature poses a challenge in linking information. This thesis aims to extract data from expanding datasets to identify and form meaningful relationships between biomedical entities. Large language models (LLMs) enable us to learn at a rapid pace. Creation of LLms from scratch are impractical. This thesis aims to collect a small dataset, containing biomedical papers, and use it to train large language models (LLMs) to extract entities from the text and learn the relationships between these entities. The experiment will be divided into two stages and utilize EU-ADR and ChemProt dataset. Starting with named entity recognition …

Improving Contextual Retrieval For Long Documents In Q & A Systems, Sakshi Garg

Master's Projects

With the vast amount of information available on the internet distributed across several lengthy documents, finding relevant information has become more important and challenging. The goal of this project is to develop advanced techniques to retrieve information from long texts in order to deliver accurate and relevant results while ensuring speed and efficiency. As part of this work, we employ techniques to address unique difficulties posed by large and complex documents. This paper presents a custom Retrieval-Augmented Generation (RAG) framework designed to improve contextual retrieval in long and multi-document settings. In this paper, we employ several techniques like summarization, semantic …

Evorgcn: Harnessing Esm-2 Evolutionary Embeddings With Relational Gcns For High-Fidelity Protein-Protein Interaction Prediction, Mohit Kunder

Master's Projects

Accurately predicting protein-protein interactions (PPIs) is essential for understanding cellular function and advancing biomedical discovery. We model PPIs as graphs, where nodes represent proteins and edges denote interactions. Using interaction data from the STRING database, we use two samples of it, namely the benchmark datasets—SH27K and SH148K—filtered by confidence score and annotated by interaction mode (multiple relations). In this project, we present EvoRGCN, a graph-based machine learning framework for PPI prediction that integrates both sequence-level (ESM-2 embeddings) and network-level information. We incorporate various Graph Neural Network architectures, including Graph Convolutional Networks, Graph Attention Networks, and Relational Graph Convolutional Networks. Our …

Synthetic Malware Generation Using Generative Ai, Phanidhar Sai Sravan Chandana

Master's Projects

Malware grows in numbers and complexity, evading conventional signature-and anomaly-based defenses and worsening extreme data sparsity and class imbalance problems for machine learning based detection. Generative models, specifically GANs conditioned on contextual embeddings like BERT have proved effective augmenting training corpora to improve classifier accuracy, but these approaches have largely produced family-specific samples In this paper, we propose a generalized augmentation scheme for generating robust malware embeddings for various families. We begin by extracting opcode sequences from 13 malware families and encoding them into three embedding methods: CountVectorizer, TF-IDF, and BERT’s ‘[CLS]‘ vectors. We therefore train standard GANs and Wasserstein …

Comparative Analysis Of Embedding Techniques With Clustering Algorithms For Malware Opcodes, Ayush Koul

Master's Projects

Malware detection and classification remain critical challenges in cybersecurity, especially as malicious software becomes increasingly sophisticated and prevalent. While much of the work involving embeddings has traditionally relied on supervised learning approaches, there is significant potential in leveraging unsupervised learning techniques to discern hidden structures in malware data. By employing embedding techniques to convert malware samples into high-dimensional vector representations, we can capture the subtle and complex patterns inherent in malicious code without relying on pre-labeled data. This unsupervised approach helps categorize malware into predefined malware families, greatly aiding in developing cybersecurity solutions. In contrast to traditional supervised models that …

Gen Ai For Malicious Network Data, Aneesh Maturu

Master's Projects

Though botnet attacks are on the rise, they also have become sophisticated and difficult to detect. Such a rising threat demands more and more sophisticated cybersecurity that leverages machine learning technology. Nevertheless, one of the biggest bottlenecks remains the unavailability of large and well-balanced datasets, particularly for malicious traffic, which hampers the efficacy of detection models. In an attempt to address this issue, our research utilizes Generative Adversarial Networks (GANs) to produce synthetic samples of botnet traffic from the CTU-13 dataset. While the majority of generative models have been targeting image data, we use GANs for a new application: generating …

Enhancing Code Review Automation With Large Language Models Using Qlora Fine-Tuning And Rags, Sumukh Naveen Aradhya

Master's Projects

In this technological era where Artificial Intelligence and Machine Learning are revolutionizing various domains, Large Language Models (LLMs) are emerging as a very powerful tool. In the software development lifecycle, it is extremely important to have reliable code reviews to ensure security and maintain code quality. This project aims to survey various existing methodologies to aid creation of efficient code review automation agents and also research on ways to make this process more efficient. Parameter Efficient Fine-Tuning (PEFT) methodologies such as LoRA and QLoRA have been explored with an additional focus on a hybrid model that combines adaptive QLoRA with …

Large Language Models For Bacterial Genomic Analysis, Manvendra Chavan

Master's Projects

Identification of bacterial gene sequences with agricultural applications has the potential to transform agricultural biotechnology. These genes can be used in environmentally friendly pest control strategies. One such use case is identifying genes with potential insecticidal properties. With an increasing number of genomic information and decreasing numbers of available annotated sequences, finding new insecticidal genes has become more challenging.The traditional methods relying on sequence alignment and annotated databases are not effective in detecting functionally relevant genes lacking close homology to known cases. This project investigates the data-driven classification of genes by sequence modeling. This research is focused on learning DNA …

Advanced Knowledge Extraction With Biomedical Data Using Llms, Akshat Krishna

Master's Projects

The rapid growth of biomedical research has led to an overwhelming volume of unstructured textual data in the scientific literature. This has necessitated the development of an automated approach for knowledge extraction and integration. In

this project, we present a comprehensive pipeline for constructing a unified biomed- ical knowledge graph by combining two well-known datasets: CHEMPROT [1],

which captures chemical–protein interactions, and EU-ADR [2], which annotates drug–gene–disease relationships. In order to identify important biomedical entities and interactions from CHEMPROT dataset, we perform Named Entity Recognition (NER) and relation Extraction (RE) using state-of-the-art biomedical models like BioBERT [3], BioGPT [4] and …

Bot Detection In Social Media Using Graphsage And Bert, Abhishek Deshmukh

Master's Projects

This project details a novel bot detection system developed to battle the ever- changing challenge of disinformation, misinformation, and other bot-generated content.

The methodology employed in this project combines the text-based analytical strength of BERT (Bidirectional Encoder Representations from Transformers) with the strength of GraphSage (Graph Sample and Aggregation) for analyzing network structures. The project concatenates BERT and GraphSage vectors to create an 896-size feature embedding with a rich blend of network and text features. This project employs a Support Vector Machine to process the concatenated embeddings, as SVM works well with high-dimensional data. This project was evaluated on two …

Ai-Enabled Anticipatory Handover Predictions In 5g Networks, Ojas Ankush Naik

Master's Projects

As users move across network cells in 5G, it is critical to maintain seamless connectivity through efficient and fast handovers. However, as 5G networks have a very dense deployment of cells and higher carrier frequencies, handovers are often more frequent and challenging, leading to failures or the ping-pong effect. In this research, we are going to use Artificial Intelligence (AI) techniques to enable predicting handover(HO) events proactively as opposed to reactively, aiming to reduce HO failures and unnecessary handovers. We develop Long-Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) models to forecast future signal measurements, predict handover trigger points, and compare …

Landslide Prediction Using Time-Series Decomposition, Reinforcement Learning-Based Feature Selection And Ml Models, Mohith Sai Venkat Ankem

Master's Projects

Landslides pose significant risks to human life, the community, and the environment, yet their prediction remains a complex and unexplored challenge. Existing prediction models often rely on surface measurements and satellite images, neglecting the critical role, in providing deeper insights into landslide analysis. The literature review highlights a lack of research in time series decomposition techniques, despite their potential to improve prediction accuracy. Similarly, feature selection methods that enhance model robustness and precision have not been adequately addressed. This study presents a novel approach to predicting landslide displacement by combining feature selection through reinforcement learning techniques with advanced time-series machine …

Quantitative Analysis Of Machine Learning Model Performance And The Need To Consider Explainability, Vishnu S. Pendyala

Open Educational Resources

This presentation, titled "Quantitative analysis of Machine Learning model performance and the need to consider explainability," delves into various metrics used for evaluating machine learning models. It thoroughly examines fundamental classification metrics like accuracy, precision, recall, and F-score, while also discussing more advanced measures such as the Kappa Statistic and Matthews Correlation Coefficient (MCC), particularly highlighting their relevance in scenarios with imbalanced datasets. The presentation underscores the importance of model accuracy in real-world applications and briefly introduces regression metrics like R-squared and F-statistic. Additionally, it addresses challenges related to data imbalance and fairness in ML models, stressing the critical need …

A Novel Reinforcement Learning Method For Efficient Cross-Training Between Real And Simulated Robots, Zhonglin Liang

Master's Theses

Training reinforcement learning policy in a simulated environment can be a go-to choice for many research topics, as a simulated environment provides flexibility and costs less than building a physical environment for training. However, policy trained in a simulated environment often fails to transfer to the real environment for problems that have a more complex dynamics. As the solution, Sim2Real was proposed and it categorizes techniques that improve the performance of the transfer from simulation to reality. Sim2Real is an area of study that focuses on utilizing simulation data to train models that can apply to real world environment. It …

Enhancing Cross-Cultural Communication In Low-Resource Language Conversational Agents, Hardi V. Trivedi

Master's Theses

Recent advancements in natural language processing (NLP) and large language models (LLMs) have facilitated the development of systems capable of generating human-like responses across a wide range of tasks. However, the majority of research has focused predominantly on English, overlooking the vast linguistic diversity globally. For true global inclusivity, extending research to other languages is crucial, particularly as it can significantly benefit various sectors such as business, healthcare, government, and education. A major challenge in this expansion is the scarcity of digital data available and the limited number of pre-trained models for low-resource languages. Our research specifically addresses these challenges …

Synthetic Data Generation For Accurate, Fair, And Private Recommender Systems, Derek B. Lilienthal

Master's Theses

A notorious challenge for recommender systems on online platforms is to accurately and fairly recommend items that align with users’ preferences while retaining user privacy. These systems often rely on historical data that is partially labeled and includes incomplete user information, leading to the over-representation of majority groups and disproportionately favoring popular items. Substituting user data with synthetic data can address these concerns, but accurately replicating real-world datasets has been a challenging problem. Recent advancements in generative AI have demonstrated the impressive capabilities of diffusion models in generating realistic data across several domains. In this work, we introduce two variations …

Controllable Physics-Informed Latent Space For Acoustic Wave Manipulation, Tristan Shah

Master's Theses

Partial Differential Equations (PDEs) are foundational descriptive tools in scientific and engineering disciplines, enabling the exploration of phenomena from material design to climate dynamics and the propagation of pandemics and wildfires. Establishing control of PDEs, especially when the underlying phenomena are only partially observable, holds great potential for fostering technological innovations. This ambition faces considerable obstacles, as a robust sample-based solution remains elusive. Addressing this challenge, this thesis introduces a framework leveraging physics-informed machine learning (ML) to enable control of PDEs. Central to this framework is an agent, equipped with sensors to perceive its environment, that generates a low-dimensional physics-informed …

An Exploration Of Dimensionality Reduction Of Dynamics On Lie Groups Via Structure-Aware Canonical Correlation Analysis, Wooyoung Chung

Master's Theses

Incorporating prior knowledge into a data-driven modeling problem can drastically improve performance, reliability, and generalization outside of the training sample. The stronger the structural properties, the more effective these improvements become. Manifolds are a powerful nonlinear generalization of Euclidean space for modeling finite dimensions. When additionally assuming that the manifold carries (Lie) group structure, this imposes a drastically stricter global constraint. The range of their applications is very wide and includes the important case of robotic tasks. We apply this idea to Canonical Correlation Analysis (CCA). In traditional CCA one constructs a hierarchical sequence of maximal correlations of up to …

Reinforcement Learning Based Planning And Control For Robotic Source Seeking Inspired By Fruit Flies, Gauravkumar Koradiya

Master's Theses

In the exploration of mobile sensing agents such as robots equipped with sensors, these agents have the potential to measure unknown scalar fields, ranging from chemical concentrations to temperature fluctuations. One of the primary interests is the robots’ ability to gravitate towards the peaks or troughs of these fields, a phenomenon termed as “source seeking”. This behavior is invaluable in practical scenarios, like detecting chemical leaks or locating survivors after catastrophes. However, the challenge becomes significantly more pronounced when operating in turbulent flow fields, where the erratic nature of odorant plumes makes predictions and simulations highly unreliable. Interestingly, fruit flies …

Domain Specific Aspect Extraction For Product Design, Divyam Sobti

Master's Theses

As technology advances, computers become increasingly proficient at interpreting and translating human language into machine-understandable text. With the help of algorithms in natural language processing (NLP), machines can now translate textual data. These algorithms help identify and extract specific text components known as aspects. The aspects represent specific attributes or topics within textual data. For instance, an Amazon review states,“ This phone has good battery life but poor camera quality,” and attributes like ‘battery life’ and ‘camera quality’ represent aspects in the text. Aspect extraction is a pivotal process involving identifying and isolating key features or topics within text. This …

Smart Robot Design And Implementation To Assist Pedestrian Road Crossing, Hovannes Kulhandjian

Mineta Transportation Institute

This research focuses on designing and developing a smart robot to assist pedestrians with road crossings. Pedestrian safety is a major concern, as highlighted by the high annual rates of fatalities and injuries. In 2020, the United States recorded 6,516 pedestrian fatalities and approximately 55,000 injuries, with children under 16 being especially vulnerable. This project aims to address this need by offering an innovative solution that prioritizes real-time detection and intelligent decision-making at intersections. Unlike existing studies that rely on traffic light infrastructure, our approach accurately identifies both vehicles and pedestrians at intersections, creating a comprehensive safety system. Our strategy …

Intellectual Property Rights And Copyright Laws In The Regime Of Artificial Intelligence (Ai) In India, Hemavathy C

Library Philosophy and Practice (e-journal)

Artificial Intelligence (AI) has been developing for two decades. The application of AI is budding quickly in business dealings, corporate communication and legal services. AI and Law Forms are increasingly important in the legal arena as they play a significant role in the economy and society. Scientists and policymakers together are facing some of the hardest problems with the advancement of machine learning, cryptology and data protection. This paper is very helpful for policymakers, economists, lawyers and technocrats in the aspect of the ethical use of AI in data protection, privacy, security and social corners turns into very relevant issues …

Data-Driven Control Of Acoustic Waves Using Movable And Flexible Scatterers, Noam Smilovich

Master's Theses

Partial Differential Equations (PDEs) serve as fundamental tools in scientific and engineering disciplines, modeling phenomena ranging from material design to climate dynamics. Developing robotic systems capable of controlling PDE-governed systems, particularly when these phenomena are only partially observable, has the potential to drive significant technological advancements. This work presents a framework that leverages physics-informed machine learning (ML) for the control of PDEs. At the core of this approach is an agent equipped with sensors that generate a low-dimensional, physics-informed representation of the environment, enabling the derivation of optimal sparse control policies for the agent’s actuators. The focus is on manipulating …

Digital Phobia: An Inquiry For Mapping The Unseen Dimension Of New Digital Anxiety, The ‘Digiphobia’, Amarjit Kumar Singh ,Library Assistant, Md. Arshad Ali , Professional Assistant, Dr. Pankaj Mathur, Deputy Librarian,

Library Philosophy and Practice (e-journal)

Background: As technology continues to advance, individuals' interactions with digital platforms have become integral to daily life. Amidst this technological evolution, a novel concern emerges—Digital Phobia, hereafter referred to as “Digiphobia.” This phenomenon, although not previously explored in scholarly literature, necessitates an in-depth investigation due to its potential impact on individuals' well-being. Our research employs a two-step methodology to investigate its existence, implications, and manifestations.

Introduction: This research paper introduces and proposes the term "Digiphobia" as a comprehensive conceptualization of anxiety arising from interactions with digital spaces, applications, and environments. The proliferation of digital technologies has led to the emergence …

Unrealvision: A Synthetic Dataset Generator For Human-Pose Estimation And Behavior Analysis, Thinh Lu

Master's Theses

For over a decade, computer vision (CV) has become an indispensable component of numerous camera surveillance applications as well as intelligent autonomous systems. Thanks to new advances in AI, Edge Computing, and IoT technologies, there is now a rapidly growing number of smart camera devices, industrial and consumer robots that are using computer vision for various applications - from human tracking and analysis, object classification, to visual inspection and anomaly detection. For most common use cases, building vision-based applications can be a straightforward and affordable task thanks to the increasing number of publicly accessible datasets and research publications. However, it …

Uncovering Weaknesses In Autonomous Driving: A Formal Approach To Adversarial Scenario Generation And Falsification, Carlos O. Hernandez

Master's Theses

Autonomous vehicles utilize advanced safety features like proactive driving assistance and pre-collision alerts to minimize the risk of accidents. However, evaluating the correct functionality of these systems is complex. First, safety systems are highly sophisticated, integrating software, networking, and hardware components, many of which rely on advanced artificial intelligence and machine learning algorithms. Second, an array of dynamic factors, including numerous actors and physical variables, can influence the performance of safety mechanisms during critical scenarios. Each actor’s unique behavior introduces unpredictability, making it difficult to anticipate future states and outcomes. This thesis presents a comprehensive framework for testing autonomous vehicle …

Resume Content Generation Using Llama 2 With Adapters, Navaneeth Sai Nidadavolu

Master's Projects

The primary objective of this project is to optimize the Llama language model to generate customized resumes containing domain-specific job descriptions and maintain the linguistic capabilities of the large language model. Building upon the prior research by Sumed Kale on Job Tailored Resume content generation using GPT-2, where he employed full fine-tuning of the model and demonstrated the capability of LLMs to generate resume content, it is evident that while effective, full fine-tuning has its limitations. Primarily, it is computationally expensive, which can pose constraints, especially for large models. Additionally, during the fine-tuning process, there is a risk of losing …

Sdebuddy - Code Documentation Using Large Language Models, Nischay Nagendra

Master's Projects

In this fast developing world of software development, it is crucial to maintain the quality of code and the developers’ productivity. This can be done effectively with good code documentation. SDEBuddy uses the latest generation of Large Language Models (LLMs) and finetuning procedures to create code documentation. In this project, state-of-the-art models such as Llama2 and Llama3 are employed to mimic the behavior of the given code and produce documentation. Such models are tuned for various programming languages and documentation formats using LoRA and QLoRA fine-tuning approaches. These models are evaluated in terms of the BLEU score, ROUGE score and …