Sameer Ambekar
Deep learning · Computer vision · 🇩🇪 🇺🇸 🇨🇦
I am pursuing my Ph.D. (Doctoral Researcher) at Technical University of Munich (TU Munich) and Helmholtz Munich in Deep learning under Prof. Dr. Julia Schnabel. Fully funded by the Konrad Zuse School of Excellence in Reliable AI (relAI) and affiliate of MCML. My Ph.D. research centres on Adapting Vision and Language models with GRPO and test-time strategies, conducted jointly at Stanford University, USA 🇺🇸 and University of British Columbia, Canada 🇨🇦.
I previously conducted a Research Stay at Stanford University, USA 🇺🇸 under Prof. Dr. Akshay Chaudhari, focusing on test-time thinking for medical imaging. Currently at University of British Columbia (UBC), Vancouver, under Prof. Dr. Xiaoxiao Li, focusing on Reasoning with Vision-Language Models.
✦ Recently, our paper Entropy Minimization without Model Collapse was selected as a Spotlight Paper [Top 3.7%] at NeurIPS 2026 main track (co-first author).
Teaching & Supervision at TU Munich
MSc Thesis & Project / IDP Supervision
Interested in Adapting foundation models, Vision and Language Models (VLMs), Fine-tuning, Reasoning, and related areas. For MSc thesis or project supervision, reach out to me directly.
Open MSc Thesis: JEPA for Medical Imaging
Looking for a motivated MSc student. Reach out directly.
View Proposal (PDF)Current Students
- • Tim Nielen, MSc Informatics, TUM Grade: 1.0 · The Dynamics of Entropy Minimization for Medical Imaging
- • Andras Gaspar, BSc, TU Munich
Seminars & Courses at TU Munich
Organizing seminar "From General to Clinical: Adapting Foundation Models for Medical Images" Slides
Organized seminar "From General to Clinical: Adapting Foundation Models for Medical Images"
News
🎉 Our paper "Entropy Minimization without Model Collapse" accepted as a Spotlight Paper [Top 3.7%] at the NeurIPS 2026 main track. Thrilled and immensely grateful! (Co-first author with Tim Nielen.)
🏆 Secured the BaCaTeC Grant funding my collaborative Ph.D. Research Stay at Stanford University, USA.
Research Stay at University of British Columbia, Vancouver 🇨🇦: Reasoning with VLMs under Prof. Dr. Xiaoxiao Li.
Research Stay at Stanford University, San Francisco 🇺🇸: Adapting VLMs to unseen tasks under Prof. Dr. Akshay Chaudhari.
Received ICML Gold Reviewer Award for paper reviewing contributions.
Hierarchical learning with Task Vectors accepted at WACV 2026 [Algorithms track].
MSc AI thesis methods accepted at WACV 2025 and CoLLAs 2024.
🏆 AwardWon Best Paper Award at MICCAIw ADSMI for Selective Test-time Adaptation.
Attended ICVSS Computer Vision summer school, Sicily, Italy.
Attended EEML 2023 Machine Learning summer school by Google DeepMind.
Research Goal & Publications
Research Goal
My research centres on leveraging deep learning to address distribution shifts, a critical challenge in real-world deployment.
Current focus: Adapting VLMs with reasoning via GRPO and test-time strategies.
Core areas: Test-time adaptation, Domain Generalization, Domain Adaptation, through variational inference, meta-learning, and surrogate model updates.
Entropy Minimization without Model Collapse: Mitigating Prediction Bias in Medical Imaging
Entropy minimization (EM) is the dominant objective for test-time adaptation, yet its failure mode (model collapse) remains poorly understood. We identify the root cause as prediction bias amplified by EM, and propose Distribution Shift Bias Reduction (DSBR), a bias-correcting objective that equalises each class's contribution to the entropy loss.
The Mean is the Mirage: Entropy-Adaptive Model Merging under Heterogeneous Domain Shifts in Medical Imaging
We demonstrate why mean merging fails under heterogeneous domain shifts and mitigate encoder–classifier mismatch by decoupling them and merging with separate adaptive coefficients.
Hierarchical Adaptive networks with Task vectors for Test-Time Adaptation
Hi-Vec uses hierarchically organised layers for dynamic test-time adaptation via automatic layer selection, cross-layer weight merging, and linear layer agreement gating, all plug-and-play on top of existing methods.
Precise Test-time Detection
Test-Time Adaptation: Non-Parametric, Backprop-free and Entirely Feedforward
GeneralizeFormer: Layer-Adaptive Model Generation across Test-Time Distribution Shifts
A lightweight meta-learned transformer generates layer parameters on the fly during inference for test-time domain generalisation, no gradient updates at test time.
Probabilistic Test-Time Generalization by Variational Neighbor-Labeling
We formulate test-time generalisation as variational inference, modelling pseudo labels as distributions to handle uncertainty and reduce error propagation.
Unsupervised Domain Adaptation for Semantic Segmentation of NIR Images through Generative Latent Search
Target-independent UDA for NIR skin segmentation by finding the nearest-clone of a target in the source domain and using it as a proxy, no target labels needed.
SKDCGN: Source-free Knowledge Distillation of Counterfactual Generative Networks using cGANs
Student 'TinyGAN' models independently learn shape, texture, and background from a pretrained BigGAN teacher without source data.
UvA DL2 course project, published 'no edits'
![[Re] Counterfactual Generative Networks](images/1/cfg.png)
Twin Augmented Architectures for Robust Classification of COVID-19 Chest X-Ray Images
Twin Augmentation boosts pre-trained deep network performance without re-training by augmenting network capacity in a paired fashion.
Education
Technical University of Munich & Helmholtz Munich
Ph.D. (Doctoral Researcher) in Deep Learning · 2023 – Present
Funded by Konrad Zuse School of Excellence in Reliable AI (relAI); affiliate of MCML. Research conducted also at Stanford University, USA 🇺🇸 and University of British Columbia (UBC), Canada 🇨🇦. Supervised by Prof. Dr. Julia Schnabel.
University of Amsterdam
Research Masters in AI (MSc AI) · Sep 2021 – Jun 2023
MSc thesis (48 ECTS, Grade: Excellent) on Test-time Adaptation for Domain Generalization: Variational Inference framework and surrogate model updates without backpropagation. Supervised by Prof. Cees Snoek, Zehao Xiao and Prof. Xiantong Zhen.
Research Experience
University of Amsterdam, AIM Lab, Research Intern
Jun 2022 – Jun 2023
Indian Institute of Technology Delhi, Research Assistant
Jan 2019 – Jul 2021
Indian Council of Medical Research (ICMR) NITM, Research Trainee
Oct 2017 – Dec 2018
DbCom Inc., New Jersey, USA, Remote Intern
Jun 2015 – Dec 2016
Scholarships & Grants
Konrad Zuse School of Excellence in Reliable AI (relAI)
DAAD-funded excellence programme training future generations of AI experts to build reliable AI systems.
BaCaTeC Grant: Stanford University Collaboration
Bavaria-California seed grant funding a Ph.D. Research Stay at Stanford University.
DigiCosme Full Master Scholarship
Université Paris-Saclay, France.
Image credits to respective organisations.
Activities & Leadership
Oxford ML Summer School (OxML 2020 & 2022), University of Oxford
PAISS 2021 ML Summer School, INRIA, Naver Labs
RegML 2021, University of Genoa
Indian Flute, recreational music
Rotaract Club of GIT, Charter Secretary & President
Website template inspired by Jon Barron.
Last updated: