Biosketch
Sameer Ambekar is a Ph.D. Student at the Technical University of Munich (TUM). He received his Masters in Artificial Intelligence (MSc AI) from the University of Amsterdam (UvA), Netherlands. For his Master’s thesis, he worked on ‘Test-Time Adaptation for Domain Generalization by generating models and labels through Variational meta-learning’ at AIM Lab, UvA. Prior to his master’s, he worked as a Research Assistant (RA) at IIT Delhi (IITD) to address Unsupervised Domain Adaptation through methods such as Variational generative latent search. He is interested in solving problems in unsupervised learning through methods such as meta-learning and variational inference alongside learning efficient and transferable features.
relAI Research
Learning to Generalize across Complex distribution shifts in Medical imaging
When test data distributions differ from training data, deep learning models often underperform, especially in medical imaging tasks like diagnosis and treatment planning. These shifts stem from scanner variability, data complexity, and limited target data access due to privacy constraints. Test-time adaptation (TTA) tackles this by optimizing source-trained models on unseen target data during inference. Unlike other methods, TTA requires no target data during training or assumptions about distribution similarity. It updates models in real-time using predictions on unlabelled test images without halting inference, enabling on-the-fly adaptation. TTA strategies include entropy minimization, self-supervision, and training modifications. Its ability to handle unseen samples and categories during inference makes it crucial for advancing medical “foundation models,” ensuring adaptability across diverse scenarios.