|
I-Con: A Unifying Framework for Representation Learning
S. Alshammari, M. Hamilton, A. Feldmann, J. Hershey, W.T. Freeman
Under review for ICLR, 2025
A unified framework that generalizes loss functions in representation learning, exposing connections across methods and achieving state-of-the-art results in unsupervised image classification on ImageNet-1K.
|
|
Vision-Language Models Do Not Understand Negation
Shaden Alshammari*, K. Alhamoud*, Y. Tian, G. Liu, P.H.S. Torr, Y. Kim, M. Ghassemi
EVAL-FoMo @ ECCV, 2024 & submitted to CVPR 2025
A benchmark evaluating negation understanding in vision-language models reveals performance limitations, with targeted improvements increasing recall by 10% and accuracy by 40%.
|
|
Long-tailed Recognition via Weight Balancing
S. Alshammari, Y. Wang, D. Ramanan, S. Kong
CVPR, 2022 (160 citations to date 🎉!)
This study explores weight balancing techniques like L2-normalization, weight decay, and MaxNorm to address bias in long-tailed recognition, achieving state-of-the-art results across five benchmarks by balancing classifier weights for rare and common classes.
|
|
Using Contact Microphones for Robot Manipulation
S. Alshammari, V. Dean, T. Hellebrekers, P. Morgado, A. Gupta
Women in Computer Vision Workshop @ NeurIPS, 2022
RoboAdapt Blind Workshop @ CoRL, 2022
This work combines visual data with contact audio to enhance manipulation in contact-rich tasks, leveraging high-frequency tactile signals from microphones to outperform single-modality approaches.
|
|
Continual Long-Tailed Recognition: Merge Tail Classes Today, Separate them Tomorrow
Y. Li, S. Alshammari, J. Jin, S. Kong
Preprint, 2022
This work introduces a continual learning approach for long-tailed recognition, using a Mean-Shift module and Supervised Contrastive loss to improve feature learning and expedite finetuning across time periods, achieving state-of-the-art performance.
|
