Research map / Label-efficient and robust learning / Self-supervised and contrastive
Unsupervised visualization of image datasets using contrastive learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Pretrain in 128D and then fine-tune in 2D, rather than training directly in 2D (ablation) | t-SimCNE · CIFAR-10 | kNN accuracy in Z (k=15): 80.0±0.2% -> 89.4±0.4% | Helped |
| Pretrain in 128D and then fine-tune in 2D, rather than training directly in 2D (ablation) | t-SimCNE · CIFAR-100 | kNN accuracy in Z (k=15), class level: 33.2±0.2% -> 51.1±0.3% | Helped |
| Pretrain with Euclidean rather than cosine loss before 2D fine-tuning (ablation) | t-SimCNE · CIFAR-10 | kNN accuracy in Z (k=15): 90.1±0.4% -> 89.4±0.4% | Mixed |
| Pretrain with Euclidean rather than cosine loss before 2D fine-tuning (ablation) | t-SimCNE · CIFAR-100 | kNN accuracy in Z (k=15), class level: 46.7±0.6% -> 51.1±0.3% | Helped |
| Pretrain with Euclidean rather than cosine loss before 2D fine-tuning (ablation) | t-SimCNE · CIFAR-100 | kNN accuracy in Z (k=15), superclass level: 65% -> 68% | Helped |
| Increase end-to-end 2D training from 1000 to 5000 epochs (ablation) | t-SimCNE · CIFAR-10 | kNN accuracy in Z (k=15): 80.0±0.2% -> 87.0±0.5% | Helped |
| Reduce the total training budget from 1500 to 1000 epochs (ablation) | t-SimCNE · CIFAR-10 | kNN accuracy in Z (k=15): 89.4±0.4% -> 88.3±0.4% | Hurt |
| Reduce the total training budget from 1500 to 500 epochs (ablation) | t-SimCNE · CIFAR-10 | kNN accuracy in Z (k=15): 89.4±0.4% -> 85.8±0.1% | Hurt |
| Use 3D cosine output instead of 128D cosine output (ablation) | SimCLR · CIFAR-10 | kNN accuracy in Z (k=15): 91.1±0.2% -> 74.0±1.4% | Hurt |
| Use 3D cosine output instead of 128D cosine output (ablation) | SimCLR · CIFAR-100 | kNN accuracy in Z (k=15), class level: 58.0±0.2% -> 15.7±0.2% | Hurt |
Closest papers by meaning
- From $t$-SNE to UMAP with contrastive learning (ICLR 2023 · similarity 0.89)
- Your Contrastive Learning Is Secretly Doing Stochastic Neighbor Embedding (ICLR 2023 · similarity 0.87)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.87)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.84)
- Single-View Graph Contrastive Learning with Soft Neighborhood Awareness (AAAI 2025 · similarity 0.83)
- FedNE: Surrogate-Assisted Federated Neighbor Embedding for Dimensionality Reduction (NeurIPS 2024 · similarity 0.83)
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.