Research map / Label-efficient and robust learning / Self-supervised and contrastive
Simplicial Embeddings in Self-Supervised Learning and Downstream Classification
Samuel Lavoie, Christos Tsirigotis, Max Schwarzer, Ankit Vani, Michael Noukhovitch, Kenji Kawaguchi, Aaron Courville
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 |
|---|---|---|---|
| Add SEM to the SSL representation | Seven SSL methods · CIFAR-100 test | linear-probe top-1 accuracy: 65.8± 0.3 -> 69.5± 0.2; 69.3± 0.3 -> 71.0± 0.3; 70.7± 0.2 -> 73.9± 0.2; 70.7± 0.3 -> 73.0± 0.2; 64.6± 0.3 -> 67.7± 0.2; | Improved |
| Add SEM to BYOL | BYOL · ImageNet in-distribution, ImageNet-V2, ImageNet-R, ImageNet-C, ImageNet-A | linear-probe top-1 accuracy: 71.9 -> 74.1; 59.2 -> 61.2; 18.8 -> 22.1; 39.5 -> 43.4; 1.65 -> 2.53 | Improved |
| Add SEM to BYOL | BYOL · ImageNet transfer: Food101, CIFAR-10, CIFAR-100, SUN, Flower (linear probe); DTD (linear probe) | linear-probe top-1 accuracy: 74.2 -> 74.7; 91.8 -> 93.5; 74.9 -> 78.6; 60.9 -> 62.1; 88.9 -> 91.5; 72.2 -> 71.9 | Mixed |
| Add SEM to BYOL | BYOL · ImageNet transfer: Food101, CIFAR-10, CIFAR-100, SUN, DTD, Flower (fine-tuned) | fine-tuned top-1 accuracy: 83.1 -> 84.7; 97.2 -> 97.2; 83.6 -> 85.6; 59.1 -> 63.3; 69.2 -> 71.3; 85.4 -> 91.7 | Mixed |
| Add SEM to BYOL | BYOL · ImageNet with 1% and 10% labels | top-1 accuracy: 51.6 -> 56.7; 67.5 -> 69.9 | Improved |
| Increase the number of simplices L while retaining softmax normalization (ablation) | BYOL + SEM · CIFAR-100 downstream classification | accuracy | Improved |
| Remove softmax normalization while increasing L (ablation) | BYOL + SEM (BYOL + Embed control) · CIFAR-100 downstream classification | accuracy | No clear effect |
| Use hard-discretized Gumbel codes during pretraining instead of SEM (ablation) | SSL representation · CIFAR-100 downstream classification | accuracy | Hurt |
| Use vector quantization instead of SEM (ablation) | SSL representation · CIFAR-100 downstream classification | accuracy | Hurt |
| Use hard argmax discretization of pretrained SEM features for downstream classification (ablation) | BYOL + SEM · CIFAR-100 downstream classification | accuracy | Improved |
| Use sparse matrix multiplication in the embedder and projector (ablation) | BYOL + SEM · CIFAR-100 | accuracy | Mixed |
| Increase simplex dimension V (ablation) | BYOL + SEM · CIFAR-100 downstream classification | accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- VicReg — VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning (ICLR 2022) · better on 1 of 1 shared result
- Barlow Twins — Barlow Twins: Self-Supervised Learning via Redundancy Reduction (ICML 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.86)
- Self-supervised Representation Learning from Random Data Projectors (ICLR 2024 · similarity 0.85)
- Measuring Self-Supervised Representation Quality for Downstream Classification Using Discriminative Features (AAAI 2024 · similarity 0.85)
- Learned Representation-Guided Diffusion Models for Large-Image Generation (CVPR 2024 · similarity 0.84)
- On the Importance of Embedding Norms in Self-Supervised Learning (ICML 2025 · similarity 0.84)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.84)
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.
