AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

UniFast-HGR: Scalable and Efficient Maximal Correlation for Multimodal Models

Hongkang Zhang, Shao-Lun Huang, Yanlong Wang, Ercan Engin KURUOGLU

ICML 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Contrastive representation theory

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add cosine reformulation to Soft-HGR baseline (ablation)UniFast-HGR · Berlinoverall accuracy (%): 65.80 -> 75.23Improved
add cosine reformulation to Soft-HGR baseline (ablation)UniFast-HGR · ImageNet-1Ktop-1 accuracy (%): 76.3 -> 77.8Improved
add variance constraints (ablation)UniFast-HGR · Berlinoverall accuracy (%): 75.23 -> 78.65Improved
add variance constraints (ablation)UniFast-HGR · ImageNet-1Ktop-1 accuracy (%): 77.8 -> 79.2Improved
add Gram-space structure-level term with diagonal retained (ablation)UniFast-HGR · Berlinoverall accuracy (%): 78.65 -> 80.62Improved
add Trivial Spectrum Suppression (diagonal removal) (ablation)UniFast-HGR · Berlinoverall accuracy (%): 80.62 -> 80.75Improved
add Trivial Spectrum Suppression (diagonal removal) (ablation)UniFast-HGR · ImageNet-1Ktop-1 accuracy (%): 79.8 -> 80.1Improved
use OptFast random-projection approximation instead of full Gram (ablation)UniFast-HGR · ImageNet-1Ktop-1 accuracy (%): 80.1 -> 79.6Mixed
reduce ImageNet local batch size from 256 to 16 (Soft-HGR) (ablation)Soft-HGR · ImageNet-1Ktop-1 accuracy (%): 76.3 -> 72.1Hurt
reduce ImageNet local batch size from 256 to 16 (UniFast-HGR with TSS)UniFast-HGR · ImageNet-1Ktop-1 accuracy (%): 80.1 -> 78.2Hurt
80% single-class imbalance in batchesUniFast-HGR vs Soft-HGR · Berlinoverall accuracy drop (points): 6.60 vs 1.85Mixed
single-label homogeneous batchesUniFast-HGR vs Soft-HGR · Berlinoverall accuracy drop (points): 13.40 vs 3.55Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy