AI Paper Atlas

Research map / Label-efficient and robust learning / Self-supervised and contrastive

Self-Supervised Classification Network

Elad Amrani, Leonid Karlinsky, Alex Bronstein

ECCV 2022 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Contrastive versus non-contrastive theory · 25 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
replace SCAN clustering-step loss with the Self-Classifier loss (ablation)SCAN · ImageNet unsupervised classificationACC: 39.9 -> 40.4Improved
use a learnable classifier instead of a fixed one (ablation)Self-Classifier · ImageNet (100 epochs)k-NN | clustering accuracy: 57.6|32.2 -> 61.7|37.3Improved
use heads with different class counts (1k+2k+4k+8k) instead of 15x1k (ablation)Self-Classifier · ImageNet (100 epochs)k-NN accuracy: 58.8 -> 61.7Improved
use multiple heads of 1k classes instead of a single 1k head (ablation)Self-Classifier · ImageNet (100 epochs)k-NN accuracy: 59.6 -> 58.6Hurt
increase batch size from 256 to 4096 (ablation)Self-Classifier · ImageNet (100 epochs)k-NN | clustering accuracy: 49.0|20.9 -> 61.7|37.3Improved
use smaller MLP (1x4096 hidden, 128 output) instead of 2x8192 hidden (ablation)Self-Classifier · ImageNet (100 epochs)k-NN accuracy: 60.0 -> 61.7Improved
set column temperature 0.05 with row temperature 0.1 instead of row 0.07 (ablation)Self-Classifier · ImageNet (100 epochs)clustering accuracy: 29.2 -> 37.3Improved
increase NN queue length from 256K to 1M (ablation)Self-Classifier · ImageNet (100 epochs)k-NN | clustering accuracy: 61.7|37.3 -> 56.8|35.5Hurt
increase NN queue length from 128K to 256K (ablation)Self-Classifier · ImageNet (100 epochs)k-NN | clustering accuracy: 59.2|36.8 -> 61.7|37.3Improved
increase training epochs from 100 to 800Self-Classifier · ImageNet unsupervised classificationACC: 37.3 -> 41.1Improved
use hidden-layer output 256 instead of 128 with 1x4096 MLP (ablation)Self-Classifier · ImageNet (100 epochs)clustering accuracy: 37.3 -> 33.5Hurt

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