Research Fellow@tirtha-debnath-57b6401a0
Tirtha Debnath
Jashore University of Science and Technology
Recent Graduate
About
Research, for me, did not begin with a grand intellectual ambition. It started with a practical question that an engineering solution could not fully answer. While building HybridFracNet-FL for my undergraduate thesis, a federated learning framework for bone fracture classification combining EfficientNetB3, DeiT-Tiny cross-attention fusion, and formal differential privacy, I kept running into the same pattern: the cross-attention fusion consistently improved generalization across federated nodes, but the theoretical reason was not obvious from the model's outputs alone. Understanding why something works, not just that it works, became the question I could not set aside. That shift in orientation is what brought me toward research, and it shapes the kind of work I want to do through this fellowship. My specific research interest is parameter-efficient adaptation of vision-language models for medical imaging, with a focus on few-shot generalization under distribution shift. Clinical imaging is a domain where labeled data is inherently scarce and where the gap between hospital sites can be large enough to make naively fine-tuned models unreliable. I am particularly interested in how lightweight adapter architectures can preserve the general representations learned during large-scale pretraining while still acquiring enough task-specific knowledge to perform well in these constrained settings. This question connects directly to my thesis work and to the broader challenge of building machine learning systems that are genuinely useful in real-world healthcare environments. I chose Md. Fahim as my supervisor because R-MMA addresses a version of this problem with unusual precision. Most adapter methods either sacrifice parameter efficiency for expressiveness or preserve pretraining quality at the cost of task-specific adaptation. R-MMA navigates this tension through a recurrent weight-sharing design where the latent token attends to the current layer's frozen features befor
Research Interests
1. arameter-efficient adaptation of vision-language models for medical image analysisspecifically investigating how recurrent weight-shared multimodal adapters can improve few-shot generalization across clinical imaging domains while preserving the pre-trained representations of large-scale models.
2. Few-shot and cross-domain generalization in vision-language models through lightweight adapter designwith a particular interest in extending recurrent multimodal architectures to specialized domains such as medical imaging where labeled data is scarce and distribution shift is significant.
3. Weight-sharing and recurrence as inductive biases for parameter-efficient transfer learning in multimodal modelsexamining how structural constraints like orthogonality regularization and modality-aware routing affect the generalization-discrimination trade-off in low-data regimes.
