Abdul Basit

Modeling network data, complex topologies, and building performance-enhanced GNNs. Developing end-to-end ML pipelines for KPI extraction, custom training loops, and distributed training on clusters.

Multimodal Text-Guided Video Segment Search

Conducted a comprehensive study on state-of-the-art techniques for text-guided temporal video retrieval. Focused on retrieving specific moments based on natural language queries by studying uniform, adaptive, and dynamic sampling strategies.

3-D Reconstruction of Scenes from Images

Analyzed state-of-the-art ML techniques for 3D reconstruction from 2D images, including NeRF, Instant NeRF, PixelNeRF, Gaussian Splatting, and Neuralangelo. Evaluated advancements in volume rendering, hash encoding, and Signed Distance Functions.

Learning Arithmetic Operations Using Small-LLM With Reinforcement Learning

This project explores the development of a lightweight Large Language Model (LLM) specialized in arithmetic reasoning, specifically focusing on integer addition and subtraction up to four digits. The project employed a two-stage training pipeline: initial Supervised Fine-Tuning (SFT) on a curriculum-based dataset to establish baseline competency, followed by a Reinforcement Learning (RL) phase using Expert Iteration (Rejection Sampling).

Event Cameras

This report investigates the fundamental operating principles and algorithmic applications of Event-based Cameras (Dynamic Vision Sensors), contrasting them with traditional frame-based acquisition. The study is divided into three primary phases: data visualization, spatiotemporal analysis, and motion estimation.

Extended Reality – Towers of Hanoi

The objective is to design and implement an immersive Extended Reality (XR) application recreating the classic "Towers of Hanoi" mathematical puzzle. The project was developed using WebXR and the Three.js JavaScript library, ensuring cross-platform compatibility and accessibility via standard web browsers on devices such as the Oculus Quest.

Continual Learning for Multi-Image Classification Task

This report presents a comprehensive evaluation of strategies for mitigating Catastrophic Forgetting in neural networks during sequential task acquisition. The project utilized the Split-MNIST benchmark to evaluate methods such as Elastic Weight Consolidation (EWC) and Experience Replay against a baseline of sequential fine-tuning.

GAN vs VAE

This paper focuses on a comparative analysis of Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), highlighting their strengths and limitations in terms of sample quality, training stability, latent space representation, and generalization. The study aims to provide insight into how different generative learning paradigms affect performance and suitability for various applications.

Style Transfer & GAN-Based Data Augmentation

Built custom CNNs to measure data augmentation impacts. Developed Conditional GANs (C-GANs) with progressive resolution growth to generate high-quality synthetic data, significantly improving CNN classifier performance. Implemented CycleGAN for unpaired image-to-image translation.

AI-Based Plant Monitoring System (Funded)

Deployed 5 different CNN architectures (ResNet50, VGG16/19, MobileNet V1/V2) on a Raspberry Pi using TF Lite for real-time disease detection. Integrated the AI system onto the end-effector of a Cable-Driven Parallel Robot operated via Python joystick libraries.

Designed PLC code for process automation, ensuring accurate feedback from sensors. Identified and resolved issues in the PLC system, and upgraded the HMI code for monitoring plant generation.