AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

On the Trade-off of Intra-/Inter-class Diversity for Supervised Pre-training

Jieyu Zhang, Bohan Wang, Zhengyu Hu, Pang Wei Koh, Alexander Ratner

NeurIPS 2023 · Label-efficient and robust learning › Recognition and long-tail · approach: Learning from unlabeled and positive-unlabeled data · 0 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
estimate class count via Extrapolation instead of Standard K=1000ImageNet subset pre-training, N=50K · CIFAR10test error rate: 29.19 -> 26.27Improved
Extrapolation instead of Standard K=1000pre-training N=50K · StanfordDogstest error rate: 64.96 -> 60.15Improved
Extrapolation instead of Standard K=1000pre-training N=50K · Stanford40test error rate: 62.45 -> 57.25Improved
Extrapolation instead of Standard K=1000pre-training N=100K · CIFAR10test error rate: 25.04 -> 23.12Improved
Extrapolation instead of Standard K=1000pre-training N=100K · StanfordDogstest error rate: 54.70 -> 53.17Improved
Extrapolation (cheap, N=5000 estimate) instead of Grid Searchpre-training N=50K · MIT67test error rate: 54.10 -> 53.01No clear effect
Extrapolation instead of Grid Searchpre-training N=100K · StanfordDogstest error rate: 51.98 -> 53.17Hurt
fewer classes than all 1000 ImageNet classes at fixed Nsupervised pre-training ResNet-18 · six downstream taskstest error rate: 29.19 -> 26.24Improved
fit theoretical error function to choose K (Fitting) vs Standardpre-training N=100K · CIFAR10test error rate: 25.04 -> 22.67Improved
predicted vs actual error from fitted functionFitting equation · CIFAR10 N=50Ktest error rate: 26.13 vs 26.25No clear effect
predicted vs actual error from fitted functionFitting equation · StanfordDogs N=100Ktest error rate: 51.69 vs 52.24Mixed
vary class count by clustering fixed samples (data process 2) instead of sampling new datasupervised pre-training · six downstream taskstest error rateImproved

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