AI Paper Atlas

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

Bayesian Optimization Meets Self-Distillation

HyunJae Lee, Heon Song, Hyeonsoo Lee, Gi-hyeon Lee, Suyeong Park, Donggeun Yoo

ICCV 2023 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives · 6 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
initialize both student and teacher from pretrained trial models instead of neither (ablation)BOSS · CIFAR-100 (VGG-16)top-1 accuracy (%): 76.48 -> 77.69Improved
initialize only the student from pretrained weights (ablation)BOSS · CIFAR-100 (VGG-16)top-1 accuracy (%): 76.48 -> 76.93Improved
initialize teacher and student from different pretrained models instead of the same onesingle-round SD, VGG-16 on CIFAR-100 · CIFAR-100 (VGG-16)top-1 accuracy (%)Improved
combine SD with BO rather than BO aloneTPE-based BO · CIFAR-100 (VGG-16)top-1 accuracy (%): 76.48 -> 77.69Improved
replace TPE with GP in BOSS (ablation)BOSS · CIFAR-100 (VGG-16)top-1 accuracy (%): 75.69 -> 77.36Improved
add SMAC variant of BOSS (ablation)SMAC BO · CIFAR-100 (VGG-16)top-1 accuracy (%): 75.93 -> 77.54Improved
use different distillation losses (KL, FitNets, Attention) inside BOSS (ablation)BOSS · CIFAR-100 (VGG-16)top-1 accuracy (%): SD 76.08 -> 77.69 (MSE); 75.63 -> 77.32 (KL)Improved
add BOSS under stronger augmentation (AutoAugment, Cutout)VGG-16 · CIFAR-100 (VGG-16, AutoAugment+Cutout)top-1 accuracy (%): 78.07 -> 80.81Improved
use K=8 candidates instead of smaller or larger KBOSS · CIFAR-100 (VGG-16)top-1 accuracy (%)Improved
vary number of warm-up trials W (best at 32)BOSS · CIFAR-100 (VGG-16)top-1 accuracy (%)No clear effect
apply SD alone or BO alone under high label noiseVGG-16 CIFAR-100 noisy labels · CIFAR-100 noisy labelstop-1 accuracy (%)Mixed

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