Research map / Label-efficient and robust learning / Self-supervised and contrastive
Contrasting Contrastive Self-Supervised Representation Learning Pipelines
Klemen Kotar, Gabriel Ilharco, Ludwig Schmidt, Kiana Ehsani, Roozbeh Mottaghi
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use a self-supervised encoder instead of a supervised ImageNet encoder | frozen ResNet50 encoder · 20 end tasks (all except ImageNet v1/v2 and Pets classification) | end-task performance | Improved |
| pre-train on ImageNet-1/4-Log (exponentially unbalanced) instead of balanced ImageNet-1/4 (ablation) | SwAV and MoCov2 encoders · 20 end tasks | average end task score: p-value = 0.0101, average end task score improvement of 1.53% | Improved |
| pre-train on ImageNet-1/2-Lin (linearly unbalanced) instead of balanced ImageNet-1/2 (ablation) | SwAV and MoCov2 encoders · 20 end tasks | end task score: p-value = 0.0777 | No clear effect |
| pre-train on a dataset similar to the end task data (ablation) | MoCov2 and SwAV encoders · SUN397, Caltech101, Kinetics action prediction, Taskonomy depth | end task performance | Improved |
| pre-train on the mixed Combination dataset instead of a dataset matched to the end task (ablation) | MoCov2 and SwAV encoders · matched end tasks | end task performance | Hurt |
| use SwAV instead of MoCov2 | ResNet50 encoders trained 200 epochs · image-level semantic tasks vs pixelwise tasks | end task performance | Mixed |
| add SwAV MultiCrop pre-processing to MoCov2 (ablation) | MoCov2 on ImageNet · 20 end tasks | end task performance | Mixed |
| use SimSiam instead of SwAV | ResNet50 encoder · 20 end tasks | normalized end task score | No clear effect |
| use Exemplar-v2 supervised training instead of vanilla supervised | supervised ImageNet encoder · 20 end tasks | end task performance | Improved |
| use self-supervised encoders on structural tasks rather than semantic tasks | self-supervised encoders vs supervised ImageNet · structural vs semantic end tasks | normalized performance over supervised | Improved |
Closest papers by meaning
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.94)
- On the duality between contrastive and non-contrastive self-supervised learning (ICLR 2023 · similarity 0.91)
- A Broad Study on the Transferability of Visual Representations With Contrastive Learning (ICCV 2021 · similarity 0.91)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.90)
- A theoretical study of inductive biases in contrastive learning (ICLR 2023 · similarity 0.90)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.88)
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.
