Ongoing Projects

Generative Adversarial Networks (GANs) have shown remarkable success in generating realistic data in fields like image synthesis, but their application to tabular data is quite limited and presents unique challenges. Tabular data, which includes structured data with categorical and numerical variables, often requires careful handling of feature distributions, dependencies, and class imbalances.

Efficacy Analysis: In this project, we aim to analyze the performance of different types of GANs in handling a wide range of imbalanced data. We plan to use different GAN architectures such as CTGAN, TGAN, or MedGAN; observe their performance, and understand their limitations. While GANs for tabular data generation offer a promising approach to overcoming the challenges of limited or sensitive data, further research is required to evaluate the efficacy of these approaches and compare them with other state-of-the-art techniques (such as SMOTE, ADASYN, and Cost-sensitive learning) used in the imbalanced domain. 

Research Gap: 

1.  While previous such works and review papers attempted to show the comparison from different papers, we aim to test all these approaches on 100+ real-world datasets with varying degrees of imbalance and dataset size, providing a systematic efficacy analysis. 

2.  Several algorithms have been proposed recently to apply GAN in tabular imbalanced data. In general, these approaches try to generate new synthetic samples to balance the dataset, without considering other associated data difficulty factors (such as overlapping, small disjunct) that complicate the learning task under constrained imbalanced datasets. I plan to investigate further and develop new algorithms to address this shortcoming.

• I am preparing to submit the initial findings at a conference.

• Codes and results will be uploaded to my GitHub repository soon.

Multiclass imbalanced classification scenarios are far more complicated than binary ones and require more careful handling. The most commonly used approach to tackle multiclass classification is the decomposition strategy:  One-vs-One (OVO) and One-vs-All (OVA). Both are widely used in real-world applications but also suffer from certain limitations (for instance, very poor performance in minority class categories).

Research Gap: A major shortcoming of the traditional approaches is that the model suffers from a loss of generalization and performance in handling minority classes.  OVR or OVO works well when the classes are more or less equal in size. However,  if the classes are skewed, applying OVR significantly increases the disparity among the number of samples, causing the model to become biased toward the majority class. Using sampling techniques may shift the bias but usually overfits the data on minority classes, resulting in poor generalization. 

Goal: In this project, a novel hierarchical decomposition framework is being developed where the classification is performed in multiple stages and the data is strategically divided, taking into consideration both the class imbalance and overlapping. This ensures better generalization and performance even in rare categories. 

Computer vision is revolutionizing medical diagnosis by automating image analysis tasks. Deep learning (DL) techniques can be utilized to process medical images such as X-rays, MRIs, CT scans, and ultrasound images with high accuracy, providing valuable insights for early detection, diagnosis, and treatment planning. Incorporating Explainable AI (XAI) tools has contributed to further enhancing the transparency and reliability of such prediction frameworks. 

In this project, we aim to apply DL approaches to detect skin diseases such as arsenic poisoning and monkeypox by analyzing dermatological images. Skin disease detection through imaging is non-invasive and can be conducted quickly with minimal discomfort for the patient. This technology can be deployed widely, from smartphones to medical-grade imaging devices, making it accessible in low-resource environments.

Around 95% test accuracy has been achieved in identifying arsenic-affected skins. I am currently testing different models and curating a larger dataset. The codes and results will be updated soon.

Braille is a tactile writing system used by people who are visually impaired. It can be read either on embossed paper or by using refreshable braille displays. It is difficult for a general person to understand a text/book that is written in Braille without having prior knowledge of the language.

The goal of this project is to develop an app that can automatically read and translate a book or text written in Bangla Braille language from images captured. I am using object detection architectures (YOLOv8) to build the recognition system. Besides, I am also curating a specialized dataset of Bangla Braille texts to support the project.