5/5 PFML achieves state-of-the-art zero shot image retrieval results on Cars-196, CUB-200-2011, and SOP, outperforming top methods by up to 7.6% Recall@1 under label noise. It’s especially effective in realistic, noisy scenarios.

4/5 Proxies in PFML help augment, not replace, sample interactions. The decaying strength of interaction helps ensure proxies stay close to the data they represent, improving alignment and feature quality, unlike prior proxy-based methods where proxies can drift away.

3/5 PFML introduces a key change: Interaction strength decays with distance. Distant samples have less influence, so outliers and mislabeled data don't dominate training. This leads to more robust learning in presence of intra-class feature diversity.

2/5 Why Potential field based Deep Metric Learning (PFML)?
Unlike triplet/proxy losses that use only a subset of interactions, PFML models all sample interactions at once using the field. PFML preserves more supervision and improves feature quality, even when using noisy labels

Excited to share our #CVPR2025 paper: "Potential Field based Deep Metric Learning "
We introduce a new DML framework that models sample interactions with a continuous potential field, with SOTA results.
🔗 shubhangb97.github.io/potential_fi...
👉 Sunday, 10:30 am Ex Hall D, Poster #431

My new paper "Deep Learning is Not So Mysterious or Different": arxiv.org/abs/2503.02113. Generalization behaviours in deep learning can be intuitively understood through a notion of soft inductive biases, and formally characterized with countable hypothesis bounds! 1/12

Excited to Present Our paper "PositiveCoOp: Rethinking Prompting Strategies for Multi-Label Recognition with Partial Annotations" at WACV 2025! @wacvconference.bsky.social 📢

🧵 A thread on what we found! 🧵

Thank you!

Hi, I would love to be added to it if possible. I am a fourth year PhD student at UIUC working on visual representation learning.