Research

2024

1. 

   Advancing Agricultural Field Segmentation Using Deep Learning

   2024-Present

   Supervisor: Dr. M. Sohel Rahman (Professor, CSE, BUET)

   Computer Vision, Machine Learning, Deep Learning

   Agricultural field segmentation using deep learning models to identify crop types, weeds, and soil. The project aims to develop a robust and efficient model for field segmentation using satellite images.
   
   Tools and Technology: Python (Pytorch), OpenCV

2. 

   Advancing Automatic Code Review and Code Refinement Generation Using Large Language Models

   2023-2024 (ArXiv)

   Supervisor: Dr. Anindya Iqbal (Professor, CSE, BUET), Dr. Toufique Ahmed (Postdoc, CS, UC Davis)

   Undergraduate Thesis, Software Engineering, Machine Learning, Natural Language Processing

   Prompting and PEFT techniques, augmenting static program metadata (function call graph) and natural language summary with human evaluation to improve code review and code refinement generation tasks
   
   Tools and Technology: Python (Pytorch), TreeSitter, OpenAI GPT API, CodeT5, CodeReviewer, CodeLlama

   Abstract

   Generating accurate code review comments remains a significant challenge due to the inherently diverse and non-unique nature of the task output. Large language models pretrained on both programming and natural language data tend to perform well in code-oriented tasks. However, large-scale pretraining is not always feasible due to its environmental impact and project-specific generalizability issues. In this work, first we fine-tune open-source Large language models (LLM) in parameter-efficient, quantized low-rank (QLoRA) fashion on consumer-grade hardware to improve review comment generation. Recent studies demonstrate the efficacy of augmenting semantic metadata information into prompts to boost performance in other code-related tasks. To explore this in code review activities, we also prompt proprietary, closed-source LLMs augmenting the input code patch with function call graphs and code summaries. Both of our strategies improve the review comment generation performance, with function call graph augmented few-shot prompting on the GPT-3.5 model surpassing the pretrained baseline by around 90% BLEU-4 score on the CodeReviewer dataset. Moreover, few-shot prompted Gemini-1.0 Pro, QLoRA fine-tuned Code Llama and Llama 3.1 models achieve competitive results (ranging from 25% to 83% performance improvement) on this task. An additional human evaluation study further validates our experimental findings, reflecting real-world developers’ perceptions of LLM-generated code review comments based on relevant qualitative metrics.
   

3. 

   Faster and Improved CD-MAWS with Suffix Automata

   2024 (Accepted at WALCOM 2025)

   Supervisor: Dr. M. Saifur Rahman (Professor, CSE, BUET)

   Suffix Automata, BioInformatics

   Introduced a refined CD-MAWS method for phylogeny estimation, significantly reducing computational complexity from max(O(mn ), O(mn log n)) to max(O(mn ), O(mnk)) while maintaining tree quality, through a revised cosine distance calculation method, binary encoding of MAWs, and the adoption of suffix automata for MAW generation.

   Abstract

   The Cosine Distance on Minimal Absent Word Sets, CD-MAWS in short, was recently introduced for alignment-free phylogeny estimation. Here we introduce a refined CD-MAWS method, significantly reducing computational complexity from O(max(km3n, km2n log n)) to O(max(m2n, kmn)) while maintaining tree quality. Here, m is the number of species, n is the size of the whole genome of a species, and k is the maximum length of a minimal absent word (MAW). This advancement is achieved through a revised cosine distance calculation method, binary encoding of MAWs, and the adoption of suffix automata for MAW generation, addressing the main computational bottleneck and setting a better runtime for alignment-free phylogenetic analysis.