Rokibul

Research Assistant

ArxEd

Dec 2025 - May 2026 • 5 mos

Research Assistant — Computer Vision / Deep Learning Optimization Designed and led a benchmarking framework to evaluate the impact of different optimization algorithms (including MTAdamV2 and MTMuon) on convolutional neural network performance for image classification tasks. Implemented and trained CNN architectures — including MobileNetV3-Large and ResNet-18 — under controlled experimental conditions to isolate the effect of optimizer choice on convergence speed, generalization, and final accuracy. Built a reproducible experimentation pipeline in Python/PyTorch to systematically compare optimizer variants across multiple model backbones, enabling fair, apples-to-apples performance evaluation. Analyzed training dynamics (loss curves, convergence behavior, stability) across optimizers to identify trade-offs relevant to real-world model selection under compute constraints. Documented methodology, results, and findings in a structured repository, contributing to the broader body of applied research on optimizer selection for CNN-based computer vision tasks. Applied skills in deep learning architecture design, hyperparameter tuning, and empirical evaluation — foundational to later work fine-tuning larger vision-language-action models (SmolVLA, OpenVLA) for robotics applications.