AI Paper Atlas

Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

Learning Clustering-based Prototypes for Compositional Zero-Shot Learning

Hongyu Qu, Jianan Wei, Xiangbo Shu, Wenguan Wang

ICLR 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Generative feature synthesis for zero-shot · 12 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ClusPro, which clusters within each attribute/object to find diverse prototypes and uses prototype-based contrastive and HSIC decorrelation losses to shape a disentangled primitive embedding space for compositional zero-shot learning.

Problem

Compositional zero-shot learning must recognize unseen attribute-object pairs, but existing methods represent each primitive with a single centroid prototype and disentangle attributes and objects only from local (batch or small subset) views, ignoring intra-primitive diversity and global data distr

Approach

Online within-primitive clustering (local-aware prototype assignment via generalized conditional gradient with a coherence regularizer, plus momentum prototype updating) discovers K prototypes per attribute and object; these prototypes anchor a prototype-based contrastive loss and an HSIC-based attribute-object decorrelation loss, and are discarded at test time.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add prototype-based contrastive learning (ablation)ClusPro without clustering baseline · UT-Zappos CWAUC: 41.0 -> 45.9Improved
Add prototype-based decorrelation learning (ablation)ClusPro without clustering baseline · UT-Zappos CWAUC: 41.0 -> 43.8Improved
Add prototype-based contrastive learning (ablation)ClusPro without clustering baseline · C-GQA CWAUC: 11.8 -> 14.2Improved
Add prototype-based decorrelation learning (ablation)ClusPro without clustering baseline · C-GQA CWAUC: 11.8 -> 13.7Improved
Use five prototypes per primitive rather than one (ablation)ClusPro · UT-Zappos CWAUC: 42.8 -> 46.6Improved
Increase prototype count beyond 5 (ablation)ClusPro · UT-Zappos CWAUC: 46.6 -> 46.0 (K=10); 46.6 -> 45.9 (K=20)Hurt
Use a high but not extreme prototype momentum coefficient (mu=0.99) (ablation)ClusPro · UT-Zappos CWAUC: 42.6 (mu=0) -> 46.6 (mu=0.99)Improved
Remove either attribute or object clustering branch (ablation)ClusPro · UT-Zappos CWAUC: 46.6 -> 44.3 (attribute only); 46.6 -> 44.1 (object only)Hurt
Replace local-aware clustering with cosine similarity updating (ablation)ClusPro · UT-Zappos CWAUC: 46.6 -> 45.0Hurt
Replace local-aware clustering with classical optimal transport (ablation)ClusPro · UT-Zappos CWAUC: 46.6 -> 45.8Hurt
Use prototype counts varying from 1 to 5 by primitive sample proportion instead of fixed K=5 (ablation)ClusPro · UT-Zappos CWAUC: 46.6 -> 46.8Improved
Use ClusPro rather than the strongest reported baselineCLIP ViT-L/14 CZSL · MIT-States CWAUC: +1.4Improved

Experimental setup

Datasets: MIT-States UT-Zappos C-GQA

Benchmarks: MIT-States UT-Zappos C-GQA

Metrics: Seen accuracy Unseen accuracy HM AUC

Baselines: CLIP CoOp PCVL CSP DFSP GIPCOL CDS-CZSL Troika PLID

Builds on: CLIP CSP DFSP Troika Sinkhorn-Knopp optimal transport

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