AI Paper Atlas

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

Exploring Hierarchical Graph Representation for Large-Scale Zero-Shot Image Classification

Kai Yi, Xiaoqian Shen, Yunhao Gou, Mohamed Elhoseiny

ECCV 2022 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Generative feature synthesis for zero-shot

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 hierarchical contrastive fine-tuning to CLIPCLIP ResNet-50 · ImageNet-21K-D (unseen classes)Hit@1: 15.22 -> 16.39Improved
use CLIP Transformer text features instead of word2vec (ablation)HGR-Net · ImageNet-21K-D (unseen classes)Hit@1: 2.35 -> 16.39Improved
add outer-loop (ancestor labels) loss (ablation)HGR-Net · ImageNet-21K-D (unseen classes)Hit@1: 15.47 -> 16.39Improved
add inner-loop (sibling negatives) loss (ablation)HGR-Net · ImageNet-21K-D (unseen classes)Hit@1: 16.16 -> 16.39Improved
use TopM negative sampling instead of random (ablation)HGR-Net · ImageNet-21K-D (unseen classes)Hit@1: 15.72 -> 16.39Improved
use TopM negative sampling instead of sibling sampling (ablation)HGR-Net · ImageNet-21K-D (unseen classes)POR: 13.46 -> 16.19Improved
use adaptive learnable layer weights instead of equal weights (ablation)HGR-Net · ImageNet-21K-D (unseen classes)Hit@1: 15.97 -> 16.39Improved
use adaptive weights instead of non-linear increasing weights (ablation)HGR-Net · ImageNet-21K-D (unseen classes)Hit@1: 15.58 -> 16.39Improved
raise outer ratio K (ablation)HGR-Net · ImageNet-21K-D (unseen classes)TOR and PORImproved
add HGR loss to CNZSL with CLIP Transformer textCNZSL(Tr) · ImageNet-21K-D (unseen classes)Hit@1: 5.77 -> 5.91Improved
add HGR loss to CNZSL with word2vecCNZSL(w2v) · ImageNet-21K-D (unseen classes)Hit@1: 1.94 -> 1.96No clear effect
separate adjacency into ancestors and descendants (ablation)DGP(Tr) · ImageNet-21K-D (unseen classes)Hit@1: 5.42 -> 5.78Improved

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