AI Paper Atlas

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

Open-Set Representation Learning Through Combinatorial Embedding

Geeho Kim, Junoh Kang, Bohyung Han

CVPR 2023 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Embedding adaptation for few-shot classification · 5 citations (OpenAlex)

Read the paper

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
Use combinatorial embeddings with multiple heterogeneous meta-classifiers instead of the GPQ retrieval baselineImage retrieval · CIFAR-10, CIFAR-100, NUS-WIDE, and CUB-200mAP: 0.108 -> 0.154 (CIFAR-100, 24 bits); 0.167 -> 0.304 (CUB-200, 24 bits)Improved
Add unsupervised pairwise relation learning to supervised meta-classification (ablation)CombEmb · CIFAR-10accuracy: 0.252 -> 0.667 (12 bits); 0.253 -> 0.692 (24 bits); 0.266 -> 0.720 (48 bits)Improved
Add consistency regularization to meta-classification (ablation)CombEmb · CIFAR-10accuracy: 0.252 -> 0.510 (12 bits); 0.253 -> 0.596 (24 bits); 0.266 -> 0.623 (48 bits)Improved
Add similarity loss to meta-classification (ablation)CombEmb · CIFAR-10accuracy: 0.252 -> 0.687 (12 bits); 0.253 -> 0.676 (24 bits); 0.266 -> 0.619 (48 bits)Improved
Add consistency loss to meta-classification plus similarity loss (ablation)CombEmb · CIFAR-10, 24 and 48 bitsaccuracy: 0.676 -> 0.692 (24 bits); 0.619 -> 0.720 (48 bits)Improved
Add consistency loss to meta-classification plus similarity loss (ablation)CombEmb · CIFAR-10, 12 bitsaccuracy: 0.687 -> 0.667Hurt
Use the proposed combinatorial-embedding pair estimator instead of k-means or RankStatsCombEmb · CIFAR-10 and NUS-WIDEmAP: 0.529 -> 0.667 (CIFAR-10, 12 bits); 0.652 -> 0.687 (NUS-WIDE, 12 bits)Improved
Increase the number of meta-classes per set from 2 to 4 (ablation)CombEmb · CIFAR-100, 72 bitsmAP: 0.179 -> 0.208Improved
Increase the number of meta-classes per set from 4 to 8 (ablation)CombEmb · CIFAR-100, 24 bitsmAP: 0.154 -> 0.141Hurt
Increase the number of meta-classes per set from 4 to 8 (ablation)CombEmb · CUB-200, 48 and 72 bitsmAP: 0.337 -> 0.337 (48 bits); 0.336 -> 0.338 (72 bits)No clear effect
Reduce the fraction of labeled examples in seen classes (ablation)CombEmb · CIFAR-10, 30% and 10% labeledmAP: 0.354, 0.572, 0.697 (30% labeled); 0.184, 0.264, 0.498 (10% labeled)Mixed
Use 48-bit representations instead of shorter bit lengths (ablation)CombEmb · CIFAR-10 and NUS-WIDE image retrievalmAP: 0.667 -> 0.720 (CIFAR-10); 0.687 -> 0.706 (NUS-WIDE)Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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