Daniel Shalam

Daniel Shalam

Ph.D. Candidate, Computer Science - University of Haifa

I am a Ph.D. candidate in Computer Science at the University of Haifa, advised by Dr. Simon Korman. My research focuses on representation learning, multimodal alignment, and generative modeling - studying how independent vision and text encoders can be aligned for few-shot multimodal learning. My work combines principles from optimal transport, flow-based methods, and data-efficient adaptation. I received my M.Sc. from the University of Haifa, where I developed optimal-transport-based frameworks for few-shot classification and self-supervised learning. I am a recipient of the Bloom Excellence Scholarship and the Dean’s List Award, and presented my research at the Israel Computer Vision Day 2024. I am broadly interested in efficient and interpretable learning systems that bridge theoretical insight with large-scale experimentation.

Publications

ECCV 2024
Unsupervised Representation Learning by Balanced Self-Attention Matching (BAM)
Daniel Shalam, Simon Korman
BAM paper image
Abstract

Balanced Self-Attention Matching (BAM) aligns self-attention representations across images to learn geometry-preserving, semantically consistent features in a completely unsupervised manner. This approach improves downstream recognition and robustness to distribution shifts.

Paper Code
ICML 2024
The Balanced-Pairwise-Affinities Feature Transform (BPA)
Daniel Shalam, Simon Korman
BPA paper image
Abstract

The Balanced-Pairwise-Affinities (BPA) transform is a training-free optimal-transport-inspired feature transform that balances pairwise relations among samples, improving transductive few-shot classification without retraining and achieving state-of-the-art results on multiple benchmarks.

Paper Code
BMVC 2023
MFSC: Matching by Few-Shot Classification
Daniel Shalam, Elie Abboud, Roee Litman, Simon Korman
MFSC paper image
Abstract

MFSC reinterprets visual correspondence as a few-shot classification task, where match candidates act as class prototypes. This approach enhances performance under low supervision and domain shifts, demonstrating competitive results on multiple correspondence benchmarks.

Paper