AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

Novel Visual Category Discovery with Dual Ranking Statistics and Mutual Knowledge Distillation

Bingchen Zhao, Kai Han

NeurIPS 2021 · Label-efficient and robust learning › Knowledge distillation · approach: Patient consistent distillation · 36 citations (OpenAlex)

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
combine global and local branches with mutual distillationNovel visual category discovery · CUB-200, Stanford-Cars, FGVC-AircraftACC: 39.5 ± 1.7% -> 47.8 ± 2.4%; 53.8 ± 2.0% -> 61.9 ± 2.5%; 66.3 ± 0.7% -> 70.4 ± 0.9%Improved
add the local branch and its part-level ranking statistics (ablation)Global-branch RankStat baseline · CUB-200ACC: 39.5 ± 1.7% -> 43.1 ± 2.3%Improved
add the BCE loss for dual ranking pseudo-label training (ablation)Ours full · CUB-200, Stanford-Cars, FGVC-Aircraft, ImageNet-100ACC: 2.2 ± 1.3% -> 47.8 ± 2.4%; 3.1 ± 0.4% -> 61.9 ± 2.5%; 5.1 ± 0.4% -> 70.4 ± 0.9%; 3.0 ± 0.3% -> 69.4 ± 2.1%Improved
add mutual-distillation sKLD loss between the branches (ablation)Ours full · CUB-200, Stanford-Cars, FGVC-Aircraft, ImageNet-100ACC: 39.8 ± 1.8% -> 47.8 ± 2.4%; 50.6 ± 2.1% -> 61.9 ± 2.5%; 60.8 ± 1.5% -> 70.4 ± 0.9%; 58.2 ± 1.2% -> 69.4 ± 2.1%Improved
add supervised cross-entropy loss on labelled data (ablation)Ours full · CUB-200, Stanford-Cars, FGVC-Aircraft, ImageNet-100ACC: 41.2 ± 2.4% -> 47.8 ± 2.4%; 52.4 ± 4.3% -> 61.9 ± 2.5%; 60.2 ± 2.7% -> 70.4 ± 0.9%; 59.1 ± 2.7% -> 69.4 ± 2.1%Improved
add consistency MSE loss across image transformations (ablation)Ours full · CUB-200, Stanford-Cars, FGVC-Aircraft, ImageNet-100ACC: 37.9 ± 4.5% -> 47.8 ± 2.4%; 50.6 ± 6.2% -> 61.9 ± 2.5%; 58.9 ± 5.7% -> 70.4 ± 0.9%; 57.2 ± 3.6% -> 69.4 ± 2.1%Improved
add self-supervised pretraining (ablation)Ours full · CUB-200, Stanford-Cars, FGVC-Aircraft, ImageNet-100ACC: 44.3 ± 3.5% -> 47.8 ± 2.4%; 58.2 ± 1.8% -> 61.9 ± 2.5%; 67.4 ± 1.3% -> 70.4 ± 0.9%; 65.3 ± 1.3% -> 69.4 ± 2.1%Improved
use two branches of the same type with mutual distillation (ablation)Global-branch configuration · CUB-200ACC: 39.5 ± 1.7% -> 41.2 ± 0.8%Improved
use two local branches with mutual distillation instead of one (ablation)Local-branch configuration · CUB-200ACC: 43.1 ± 0.9% -> 44.7 ± 1.1%Improved
use global and local branches rather than two local branches (ablation)Two-branch design · CUB-200, Stanford-Cars, ImageNet-100ACC: 44.7 ± 1.1% -> 47.8 ± 2.4%; 57.9 ± 0.5% -> 61.9 ± 2.5%; 65.7 ± 1.4% -> 69.4 ± 2.1%Improved
increase global/local feature-bank size from 2048 to 32768 (ablation)Ours full · CIFAR-10, CUB-200, Stanford-Cars, ImageNet-100ACC: 91.6% -> 93.0%; 47.8% -> 48.4%; 61.9% -> 62.9%; 69.4% -> 71.3%Improved
increase local part dictionary V size from 2048 to 32768 (ablation)Ours full · CIFAR-10, CUB-200, Stanford-Cars, ImageNet-100ACC: 91.6% -> 92.0%; 47.8% -> 49.3%; 61.9% -> 63.7%; 69.4% -> 71.4%Improved

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