Naimul Haque

Experiences

Teaching Experiences

TA at Alfred University
Fall 2024
Level: Undergraduate
• Python for Power Systems - 19121 - ELEC 330 - 01

Assistant Professor at Uttara University
August 2023 - Present (Full Time)
Conducting undergraduate classes
• Structured Programming[60I][60J][Fall 23]
• Discrete Mathematics[60F-EVE][Fall 23]
• Data Structure[57B][Fall 23]

Assistant Professor at Department of CSE, Manarat International University
Mar 2023 - August 2023 (Full-Time) Conducting undergraduate classes
• CSE-437: Computer Vision and Robotics[Fall 23]
• CSE-433: Neural Network and Fuzzy Systems[Fall 23]
• CSE-210: Algorithm[Fall 23]
• CSE-211: Assembly Language[Fall 23]

Lecturer at CSE, Manarat International University
April 2019 - Mar 2023 (Full-Time)
Conducting undergraduate classes
• CSE-411: Artificial Intelligence [Spring 23]
• CSE-437: Computer Vision and Robotics[Fall 21][Spring 22][Fall 22]
• CSE-433: Neural Network and Fuzzy Systems[Summer 20][Summer 20][Spring 21][Spring 22][Summer 22]
• CSE-436: Pattern Recognition[Summer 20][Summer 21]
• CSE-210: Algorithm[Fall 19][Fall 20][Summer 21][Fall 21][Spring 22][Summer 22][Fall 22]
• CSE-104: Discrete Mathematics[Fall 21][Spring 22]
• CSE-314: Microprocessor and Interfacing [Sum 2019][Spring 20][Sum 20][Sum 20][Spring 21][Sum 21][Spring 22][Fall 22]
• CSE-410: Computer Networking[Summer 19]
• EEE-101: Electrical Circuits[Summer 19][Fall 19][Summer 20]
• CSE-313: System Analysis and Design[Fall 19]
• CSE-312: Software Development[Spring 20]
• CSE-414: Internet and Web Development[Spring 20][Summer 20][Fall 20]
• CSE-402: Software Engineering[Spring 20][Summer 20][Fall 20]
• CSE-211: Assembly Language[Summer 20][Fall 20][Spring 21][Fall 22]
• PHY-102: Physics[Fall 21]
• CSE-102: Structured Programming[Summer 22][Fall 22]

Adjunct Lecturer at Department of CSE, Ahsanullah University of Science and Technology
February 2020 - June 2020 (Contract) Department of CSE, Ahsanullah University of Science and Technology
· Conducting undergraduate Lab class: CSE 3214: Operating System.

Research Experiences

Graduate Research Assistant at Alfred University
January 2024 - Present
At Alfred University, I am conducting advanced research in AI and Deep Learning, focusing on catastrophic forgetting in Large Language Models (LLMs). My work involves evaluating LLMs for sequential fine-tuning on natural language understanding tasks and developing techniques to mitigate performance degradation across tasks.

Additionally, I have been involved in setting up the Advanced Power Distribution Management Lab, which includes the installation and debugging of GE's Advanced Distribution Management System (ADMS) and Automated Energy Management System (AEMS). I resolved compatibility issues and documented the installation process for future reference. I also optimized lab setups to improve the research and learning environment for students working on AI applications in the power industry.

Lab Director at Cognition.AI
November 2020 - 2022
Cognition.AI is my Research Lab where I supervise students for their undergraduate thesis on Computer Vision and Natural Language Processing using Deep Learning.

Programming Experiences

Indie Game Developer at Unity Game Engine
November 2020 - 2022
I developed few indie games using Unity Game Engine. Lost is a platformer based 2D game developed using Unity and C#. Android APK is available, download, play and enjoy.

Education

M.Sc. in Electrical Engineering Alfred University
2024
Thesis: Catastrophic Forgetting for Continual Instruction Fine-tunning LLMs, Supervisor: Dr. Junpeng Zhan Professor, Renewable Energy Engineering
CGPA: 4.0/4

M.Sc. in Computer Science Jahangirnagar University
2019 - 2022
Thesis: Bangla License Plate Restoration using a Super Resolution Generative Adversarial Networks, Supervisor: Dr. Mohammad Shorif Uddin Professor, Department of CSE (JU)
CGPA: 3.53/4

B.Sc. in Computer Science and Engineering Ahsanullah University of Science and Technology
2014 - 2018
Thesis: Grayscale Image Colorization, Supervisor: Dr. S.M.A. Al-Mamun P, Professor \& Dean, Department of CSE (AUST)
CGPA: 3.712/4

Publications

Published

Bangla Toxic Comment Classification and Severity Measure Using Deep Learning
Naimul Haque, Md. Bodrul, Abdullah, Towfiq, Mehorab
Accepted in - International Conference on Recent Progresses in Science, Engineering and Technology (ICRPSET-2022) (Conference Paper)
DOI: 10.1109/ICRPSET57982.2022.10188551

Automatic Bangla License Plate Recognition System for Low-Resolution Images
Naimul Haque, Samira Islam, Rabeiya Afrose, Mohammad Shorif Uddin
Accepted in - 2022 4th International Conference on Sustainable Technologies for Industry 4.0 (STI), 17-18 December, Dhaka (Conference Paper)
DOI: 10.1109/STI56238.2022.10103289

Grayscale Portrait Colorization using CNNs and Pretrained VGG-Face Descriptor
Naimul Haque, Samin Shahriar Tokey (co-supervisor)
Accepted in - 2019 22nd International Conference on Computer and Information Technology (ICCIT), 17-18 December, Dhaka (Conference Paper)
DOI: 10.1109/ICCIT48885.2019.9038376

Grayscale Portrait Colorization using Optimization and MTCNN Face
Naimul Haque, Samin Shahriar Tokey
Published in - Southeast University Journal of Computing Sciences, Volume 1, No. 1, June 2021

Pre-Prints

FaceGemma: Enhancing Image Captioning with Facial Attributes for Portrait Images
Naimul Haque, Iffat Labiba, Sadia Akter
Submitted to - International Conference on Pattern Recognition (ICPR), Kolkata
arXiv:2309.13601v1 [cs.CV] [Round 1]

BdSpell: A YOLO-based Real-time Finger Spelling System for Bangla Sign Language
Naimul Haque, Meraj Serker, Tariq Bin Bashar
Accepted in - ICRIC 2023 International Conference on Recent Innovations in Computing, Central University, 26-27 October, 2023
arXiv:2309.13676v1 [cs.CV]

Face-Att: Enhancing Image Captioning with Facial Attributes for Portrait Images
Naimul Haque, Iffat Labiba
Submitted in - 2nd International Conference on Information and Communication Technology For Sustainable Development, 14-16 September, 2023, Bangladesh University of Professionals (Conference Paper)
Read Source Code

Catastrophic Forgetting in LLMs: A Comparative Analysis Across Language Tasks
Naimul Haque
Submitted in - ACL ARR 2024 (October submission)
Link [Submitted]

Certifications

Machine Vision in the Electric Industry with Deep Learning (2024)
EPRI | 22 Professional Development Hours (PDHs)

• Tools Used: YOLOv8 for object detection, Colab, TensorFlow for deep learning applications.
• Methods Applied: Deep learning fundamentals, object detection using pre-trained models, practical implementation of machine vision in the electric industry.
• Key Learnings: Gained hands-on experience in applying machine vision techniques to solve industry-specific challenges in the electric sector, specifically using deep learning for efficient power grid management and infrastructure monitoring.

Projects

Research Projects

Improving Continual Fine-Tuning with Knowledge Distillation
In this project, I applied knowledge distillation techniques during the fine-tuning of LLMs for SST-2, MRPC, and CoLA tasks (GLUE benchmark). Knowledge distillation helps transfer knowledge from a larger teacher model to a smaller student model, enhancing learning across tasks while reducing catastrophic forgetting.
Results:

• Llama-3.2-3B* showed significant learning improvements due to the distillation process.
• Llama-3.2-3B displayed more stability across tasks with lower forgetting.

This approach holds promise for enhancing continual learning capabilities in language models.

Backdoor Attacks on Fine-Tuned Llama-LLM for Twitter Sentiment Analysis
This project focuses on studying backdoor attacks on a fine-tuned Llama-LLM model used for Twitter sentiment analysis. The project code is provided in the "alpaca-fine-tuning twitter sentiment 1.6M.ipynb" Jupyter notebook. The goal is to investigate the potential vulnerabilities of the model to backdoor attacks and explore the use of the Alpaca LoRa framework for model quantization. Model Weights available @ HuggingFace HuggingFace.

Real-Time Bangla Finger Spelling for Sign Language
The Real-Time Bangla Digit and Alphabet Detection project aims to develop an advanced computer vision system capable of accurately detecting digits and alphabets in Bangla sign language. Leveraging the power of YOLOv5, a state-of-the-art object detection algorithm, this project will enable real-time recognition of Bangla sign language gestures for digits and alphabets. The system's ultimate goal is to utilize these real-time detections to spell out words, enhancing communication and facilitating meaningful interactions for individuals with hearing and speech impairments.
Training and Dataset Available @ RoboFlow Roboflow.

Enhancing Face Captioning using Facial Attributes
The goal of this project is to generate accurate and descriptive captions for facial images using a combination of facial attributes and natural language processing techniques. By incorporating facial attributes, the model can generate more contextually relevant captions for each image.
Dataset Available @ Zenodo Zenodo.

Competitions

Google – AI Assistants for Data Tasks with Gemma
Rank: Unranked
My Notebook
Participated in developing AI assistants for automating data-related tasks using multi-agent systems.

LLM 20 Questions
Rank: Unranked
My Notebook
Focused on prompt engineering techniques with LLaMA 3.1 for solving a series of 20 questions using reflective AI methods.

Jigsaw Rate Severity of Toxic Comments
Rank: 1228/2301
Score: 0.88
Team Members: naimul314, athtowfiq, mdbodrulalam, mehorab
Ranked relative ratings of toxicity between comments in a code competition with 2301 teams, held 3 years ago.

Published Models

Fine-Tuned Models

GlueOrca
Naimul Haque
Fine-tuned on - GLUE Benchmark Tasks: SST-2 (Sentiment Analysis), MRPC (Paraphrase Detection), CoLA (Linguistic Acceptability), and MNLI (Natural Language Inference)
Performance Highlights:

• Pretrained Performance: 0.61
• Forgetting Rate: 0.35
• Learning Capability: 1.05
• Training Performance: 0.75

GlueQwen
Naimul Haque
Fine-tuned on - GLUE Benchmark Tasks: SST-2 (Sentiment Analysis), MRPC (Paraphrase Detection), CoLA (Linguistic Acceptability), and MNLI (Natural Language Inference)
Performance Highlights:

• Pretrained Performance: 0.59
• Forgetting Rate: 0.42
• Learning Capability: 0.97
• Training Performance: 0.73

Datasets

Datasets

Bangla License Plate Dataset 2.5k
Naimul Haque, Mirajul Islam, Afrose Tithy, Samira Islam
Published in - Zenodo, September 24, 2022 (CV-Dataset)

FaceAttDB: A Multilingual Dataset for Facial Attribute Captioning
Naimul Haque, Abida Sultana
Published in - Zenodo, June, 2023 (CV-NLP-Dataset)

nBangla Toxic Comments DB
Naimul Haque, Badrul
Published in - HuggingFace, June, 2023 (NLP-Dataset)