Ibna Kowsar

Research Associate, CIDA Lab, Tennessee State University
Upcoming Ph.D. Student, UC Davis (Fall 2025)
Email: kawser.ibn.93@gmail.com || ikowsar@my.tnstate.edu

I'm currently a Research Associate at CIDA Lab, Tennessee State University, and will begin my Ph.D. in Computer Science at UC Davis in Fall 2025.

My research interests include:

  • Deep Learning
  • Model Interpretability
  • Large Language Models
  • Cross-Domain Transfer Learning
  • Electronic Health Records Analysis
  • Self-supervised & Unsupervised Methods

Experience

Research Associate

CIDA Lab, TSU

Advancing deep learning and transfer learning techniques for medical imaging and EHR data.

July 2025 - Present

Graduate Research Assistant

CIDA Lab, TSU
  • Implemented and enhanced deep learning algorithms for image and tabular data.
  • Investigated domain adaptation strategies to improve model generalization.
August 2023 - June 2025

Machine Learning Engineer

Apurba Technologies Ltd.
  • Implemented and improved existing ML systems for OCR and document analysis.
  • Researched adversarial attacks and layout analysis for Bengali OCR.
July 2021 - January 2023

Lecturer

BRAC University
  • Assessed and mentored over 150 students per term across core CS courses.
  • Delivered lectures and developed course materials for CSE110, CSE221, CSE260, CSE321, CSE370, and CSE461.
October 2021 - August 2023

Teaching Assistant (Undergraduate)

BRAC University
  • CSE110: Programming Language I - Lab instructor and student consultations.
  • CSE111: Programming Language II - Video tutorials and assignment grading.
January 2020 - May 2021

Education

BRAC University

Bachelor of Science
Computer Science & Engineering

CGPA: 3.87/4.00 (Highest Distinction)

Thesis Title: Facial Expression Recognition: Convolutional Attentional Masking Network and Ensemble Approach

Abstract

Facial expression plays a significant role in human communication. The necessity of recognizing facial expression is increasing rapidly as it can be implemented in various important fields such as in human-computer interactions, medical care, autonomous transportation systems etc. The facial expression detection has been accomplished by the analysis of convolutional neural networks on the micromotors and action units. In this thesis, we have introduced a new variant of residual architecture named CAMnet which uses the split attentional module and the masking module mechanisms simultaneously. Also, the model performs better compared to other models without using any pretrained weights on small dataset like FER2013. Additionally, along with the CAMnet an ensemble model has been implemented and we have achieved 76.12% accuracy on the FER2013 test set.

2017 - 2021

Publications

Attention-based Imputation of Missing Values in Electronic Health Records Tabular Data

Ibna Kowsar, Shourav B. Rabbani, Manar D. Samad

International Conference on Pattern Recognition (ICHI 2024)

Paper
Abstract

The imputation of missing values (IMV) in electronic health records tabular data is crucial to enable machine learning for patient-specific predictive modeling. While IMV methods are developed in biostatistics and recently in machine learning, deep learning-based solutions have shown limited success in learning tabular data. This paper proposes a novel attention-based missing value imputation framework that learns to reconstruct data with missing values leveraging between-feature (self-attention) or between-sample attentions. We adopt data manipulation methods used in contrastive learning to improve the generalization of the trained imputation model. The proposed self-attention imputation method outperforms state-of-the-art statistical and machine learning-based (decision-tree) imputation methods, reducing the normalized root mean squared error by 18.4% to 74.7% on five tabular data sets and 52.6% to 82.6% on two electronic health records data sets. The proposed attention-based missing value imputation method shows superior performance across a wide range of missingness (10% to 50%) when the values are missing completely at random.

• Accepted in ICHI2024 [will be online soon]

A Deep Learning Based Unified Solution for Character Recognition

Avishek Das, AKM Shahariar Azad Rabby, Ibna Kowsar, Fuad Rahman

International Conference on Pattern Recognition (ICPR 2022)

Paper
Abstract

Optical Character Recognition(OCR) has become a crucial area of research due to the vast number of digitized documents to lessen the dependency on paper. One can save time and money on data entry by automatically extracting information off paper and putting it where it needs to go. There has been much research on OCR systems for different languages, but a unified system that is agnostic to language does not exist. In this work, we propose a multi-headed resunet++ based solution that can recognize the low resource languages(Bangla, Assamese, etc.) and performs well on resource-rich languages(such as English, Arabic, etc.). The backbone of the solution, i.e., resunet++, is fundamentally designed for medical image segmentation that is very complex. As the low representative languages are mostly of cursive style and complex in nature, this backbone can help share those higher-level features and pass them to the lower level. Our proposed solution is applied to isolated characters of Bangla, Assamese, and English languages. For Bangla, the segmentation is done by our developed method, and the dataset was pre-segmented for the other two languages. Applying the solution, we achieved a satisfactory performance.

Towards building a Bangla text recognition solution with a Multi-Headed CNN architecture

Md. Majedul Islam, Avishek Das, Ibna Kowsar, A K M Shahariar Azad Rabby, Nazmul Hasan, Fuad Rahman

IEEE International Conference on Big Data (IEEE BigData2021)

Paper
Abstract

Bangla is among the ten most popular languages in the world by the number of speakers. The task of Bangla recognition is quite challenging than other languages because of the existence of graphemes of multiple single characters, and diacritics of vowels and consonants. The purpose of this study is to develop an innovative large-scale Bangla OCR solution based on character-level recognition. Two types of documents were used to test our method: handwritten and printed. In addition, our method was applied to the handwritten documents as well as three subdomains of the printed domain: computer-composed, letterpress, and typewritten documents using our proposed attentionbased multi-headed CNN architecture. Extensive testing shows that our method provides state-of-the-art performance on both handwritten and printed texts.

• State-of-the-art performance on both handwritten and printed texts

DOI: 10.1109/BigData52589.2021.9671653

A Novel Approach to Enhance Safety on Drowsy Driving in Self-Driving Car

Md. Motaharul Islam; Ibna Kowsar ; Mashfiq Shahriar Zaman; Md. Fahmidur Rahman Sakib; Nazmus Saquib. (Springer Nature)

Paper
Abstract

Drowsy driving centric accidents are increasing at a frightening rate. Needless to say that the state-of-the-art technologies only have competencies in detecting drowsiness and alerting the drowsy driver. Existing methods have some remarkable hindrances in the domain of handling the distressed situation. Therefore these methodologies are ineffective to take additional safety measures if the driver is not proficient enough to operate the vehicle even though an alarm is given. Consequently, after evaluating the existing methodologies and the growth of autonomous vehicles, we have proposed an innovative approach that detects driver drowsiness in real-time. Our suggested model can locate a nearest available safe parking space and reach the parking location after initiating the autonomous driving mode to ensure safety. The proposed methodology has achieved an accuracy of 98%.

• State-of-the-art approach to handle post-alarm condition on autonomous vehicle

An Algorithmic Approach to Driver Drowsiness Detection for Ensuring Safety in an Autonomous Car

Md. Motaharul Islam; Ibna Kowsar ; Mashfiq Shahriar Zaman; Md. Fahmidur Rahman Sakib; Nazmus Saquib

Paper
Abstract

Human-centric accidents are increasing gradually and one of the dominant causes of the accidents is driver drowsiness. Therefore, to lessen the accidents related to drowsiness, methods that are capable of observing facial expression to detect drowsiness have been proposed by researchers in recent decades to ensure safety. However, the state-of-the-art models only have the competency in determining the drowsiness and alarming the driver. Traditional approaches divide the detection method into two stages, such as detecting drowsiness from the driver's facial features and further apprising the driver. Hence, the existing models are inadequate to take any additional safety procedures to ensure more safety if the driver remains unable to operate the vehicle after giving an alarm. Analyzing these approaches and because of the increasing reliance on the vehicles, we have introduced an algorithmic approach in which the proposed system can locate a safe parking space after the determination of drowsiness and can also deliver a distress message to the authority informing about the situation while reaching at the safe parking space to assure safety from the incompetent, drowsy driver.

An efficient Metaheuristic Approach for Finding Motifs from DNA Sequences

Syed Md. Shamsul Alam, Ibna Kowsar, Md. Al-Junaed Islam, Shurid Shahriar Zaman, Faisal Bin Ashraf

Paper
Abstract

Finding patterns of the short sequences in DNA, RNA protein sequence has immense biological significance. The characterization and recognition of motifs is therefore an important method for a more in-depth understanding of genes or proteins in their structure, function and relations of evolution. This is one of the classical problems in the field of computational biology and which is an NP Hard problem. In this paper, we have proposed an evolutionary approach to get the motifs from DNA sequence by searching candidate motifs using heuristic way from the data. We have included various mutation techniques in an evolutionary approach and found an efficient way to calculate the fitness of our candidate motifs. We have evaluated the fitness of found motifs from our approach with benchmark data sets. Our method performs better results in terms of accuracy and specificity.

Skills

Programming Languages & Tools
Workflow

Awards & Certifications

Projects

Facial Expression Recognition

Paper Code

• Proposed a state-of-the-art Convolutional Attentional Masking Network and Ensemble Network.

• Obtained 76.12% accuracy on detecting FER

CNN, State-of-the-art

Driver Drowsiness Detection and Alarming System

Paper

• Real-Time driver observation using OpenCV & python API

• State-of-the-art approach to handle post-alarm condition on autonomous vehicle

• The system can book a local parking space and reach there if the driver is sleepy

Python, Opencv, ML

SimoBot

Code Simulation