Hi, I'm Farhan, an incoming PhD student at KSoC, UofU. I'm currently working as a lecturer at IUT-CSE and a part-time research intern at Yaana Technologies. Prior to that, I obtained my Bachelor's degree from IUT-CSE.

Currently, I am exploring domain adaptation and continual learning for Vision-Language Models (VLMs), specifically its application in computer use. Broadly speaking, my research interests span:

• Continual & Lifelong learning for VLMs
• Agent-centric Benchmarks
• Multimodal Adaptation & Transfer Learning

News and Updates

• Jan 25: BanTH, got accepted in Findings of NAACL, 2025!
• Dec 24: BnSentMix, got accepted in LoResLM at COLING 2025!
• Oct 24: Visual Robustness Benchmark for VQA is accepted at WACV 2025. I’m grateful to my wonderful teammates and advisors at IUT-CSE.
• Oct 24: Check out our pre-prints on Bangla VQA system, ChitroJera, and multi-label hate speech detection on transliterated Bangla, BanTH.
• Oct 24: FourierKAN outperforms MLP on Text Classification Head Fine-tuning got accepted in FITML at NeurIPS 2024. Shoutout to my teammate Abdullah Al Imran.
• Sep 24: Our work on Bangla back-transliteration, BanglaTLit got accepted in Findings of EMNLP, 2024. Great work from team Penta!
• Aug 24: Check out the preprints on Fourier-KAN Text Classification and BnSentMix: Code-Mixed Bengali Sentiment Analysis, available on arXiv.
• Jul 24: New preprint on my undergrad thesis, Visual Robustness Benchmark for VQA, is available on arXiv.
• May 24: Finalists at Robi Datathon 3.0, Bangladesh’s largest data analysis event with 3,500+ participants. Another competition with team Penta!
• May 24: Participated in the EXIST-2024 shared task with my amazing team from Penta Global.
• Jan 24: Our VQA Survey got accepted in Information Fusion.
  

Research Highlights

Visual Robustness Benchmark for Visual Question Answering (VQA)

Md Farhan Ishmam*, Ishmam Tashdeed*, Talukder Asir Saadat*, Md Hamjajul Ashmafee, Abu Raihan Mostofa Kamal, Md Azam Hossain

• Can VQA models maintain their performance in real-world scenarios, especially when prone to realistic visual corruptions, e.g. noise, blur?
• How can we evaluate the robustness of VQA models when subjected to such common corruptions?
• We propose a large-scale benchmark and robustness evaluation metrics to evaluate VQA models before deployment.
• We quantify aspects of performance drop, e.g. rate, range, mean. We found that models with higher accuracy are not necessarily more robust.

BanglaTLit: A Benchmark Dataset for Back-Transliteration of Romanized Bangla

Md Fahim*, Fariha Tanjim Shifat*, Md Farhan Ishmam*, Deeparghya Dutta Barua, Fabiha Haider, Md Sakib Ul Rahman Sourove, Farhad Alam Bhuiyan

• How can we enhance the representation of Romanized Bangla for automatic back-transliteration using seq2seq models?
• We propose large-scale pre-training corpus and Bangla back-transliteration datasets for fully fine-tuning language encoders and seq2seq models.
• We aggregate representations from transliterated Bangla encoders with seq2seq models [Dual Encoders->Aggregation->Decoder architecture] to achieve SOTA on Bangla back-transliteration.

From Image to Language: A Critical Analysis of Visual Question Answering (VQA) Approaches, Challenges, and Opportunities

Md Farhan Ishmam, Md Sakib Hossain Shovon, Muhammad Firoz Mridha, Nilanjan Dey

• Comprehensive survey on VQA datasets, methods, metrics, challenges, and research opportunities.
• New taxonomy that systematically categorizes VQA literature and multimodal learning tasks.
• Novel real-world applications of VQA in domains e.g. assistive technology, education, and healthcare.