Dipam Goswami

I am a postdoctoral researcher in the Learning and Machine Perception (LAMP) group at the Computer Vision Center, Barcelona. I work on Vision-Language Models (VLMs) and Multimodal Large Language Models (MLLMs), with a broader interest in understanding and improving the representations of large foundation models. I also work on Continual Learning which studies how deep learning models can keep learning from new data over time without forgetting what they already know, with a particular focus on understanding and controlling how their embedding spaces change as they learn.

Recent research

My recent works focuses on multimodal and generative models:

• Vision-Language Models: studying cross-modal alignment in contrastive VLMs and exploring training-free approaches.
  • Improving intra-modal alignment in CLIP by by decomposing CLIP’s projection layers (IsoCLIP - CVPR 2026).
  • Improving few-shot image classification using cross-modal prototypes exploiting a task-semantic image subspace (Project and Mix - Preprint).
• Continual Learning with VLMs: a survey and taxonomy beyond forgetting (Preprint) and exploiting the semantic knowledge of pre-trained text-encoders (Preprint).
• Data Attribution with Diffusion Models: mirrored unlearning approach for training data attribution, i.e. tracing which training data influences a generated output (MUCS - NeurIPS 2026).

PhD research

My PhD focused on how models learn when data is not available all at once: arriving over time or spread across clients. I developed efficient methods that work without storing old data, without re-training, or with minimal communication, by studying how embedding spaces change during learning:

• Continual Learning: exemplar-free class-incremental learning (FeCAM - NeurIPS 2023), few-shot class-incremental learning (CVPRW 2024) and compensating for feature drift (ADC - CVPR 2024, LDC - ECCV 2024).
• Federated Learning: training-free, communication-efficient methods built on pre-trained models (FedCOF - NeurIPS 2025).
• Information Retrieval: continual learning of dense retrieval embedding models that stay compatible with existing embedding indexes (QDC - CoLLAs 2025).

Before starting my PhD, I worked on continual learning for object detection (ICCV 2023) and semantic segmentation (WACV 2023), defect detection in SEM images (SPIE Advanced Lithography 2022), cell detection from urine microscopic images (ISBI 2023, UMID dataset) and graph-based approaches for generation of dimensioned floorplans (AI EDAM 2021, SCCE 2024).

Background

I completed my PhD in April 2026 at the Computer Vision Center, Universitat Autònoma de Barcelona, supervised by Joost van de Weijer and Bartłomiej Twardowski, with the thesis Understanding the Embedding Space in Continual and Federated Learning. Before that, I received a B.E. in Computer Science and an M.Sc. in Mathematics from BITS Pilani, India in 2022.

Research visits and industry experience

• Sony AI Barcelona (2025): research internship on training data attribution for Diffusion Models with Joan Serrà.
• KU Leuven, PSI group Belgium (2025): research stay on Vision-Language Models with Tinne Tuytelaars and Gido van de Ven.
• IDEAS NCBR Warsaw (2024): research stay on continual learning of dense retrieval embedding models with Bartłomiej Twardowski.
• DFKI Kaiserslautern, Germany (2022): research assistant in the Augmented Vision Group under Didier Stricker where I worked on continual learning for semantic segmentation.

Recognition and talks

• Invited talk on Understanding the Embedding Space in Continual, Federated and Multimodal Learning at the MICC seminar, University of Florence (2026).
• Invited talks on Exemplar‑Free Continual Learning at the GMUM seminar, Jagiellonian University, Kraków (2024); at the Data Science Summit, Warsaw (2024), at the CVML reading sessions at the University of Barcelona (2024) and at the Deep Learning Barcelona Symposium (2023).
• Top reviewer at NeurIPS 2024 and outstanding reviewer at BMVC 2024.
• Innovation award (team “Continual Learners”) at the Continual Test-time Adaptation challenge, Visual Continual Learning Workshop, ICCV 2023, where I also gave an oral presentation.

Community service

• I served as a reviewer for ICML, ICLR, NeurIPS, AISTATS, CoLLAs, CVPR, ICCV, ECCV, WACV and BMVC, and for journals including IEEE TPAMI, IJCV, TMLR, IEEE TNNLS and IEEE TIP.
• I co-organize the CoLLAs Seminars, a monthly online seminar series on lifelong learning launched in May 2026.

I am open to postdoctoral and research scientist positions starting in January 2027. Feel free to reach out.