Research map / Label-efficient and robust learning / Self-supervised and contrastive
The Mechanism of Prediction Head in Non-contrastive Self-supervised 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 |
|---|---|---|---|
| Train off-diagonal entries of an identity-initialized prediction head | Non-contrastive self-supervised learning · Two-feature theoretical setting | inter-neuron correlation: |\mathbf{Corr}(f_1(X),f_2(X))| \leq O(\frac{1}{\sqrt{d}}) versus |\mathbf{Corr}(f_i(X),f_j(X))| \geq 1 - O(\frac{1}{\sqr | Helped |
| Keep the prediction head fixed at identity instead of training its off-diagonal entries (ablation) | Non-contrastive self-supervised learning · Two-feature theoretical setting | inter-neuron correlation: |\mathbf{Corr}(f_i(X),f_j(X))| \geq 1 - O(\frac{1}{\sqrt{d}}) | Hurt |
| Add an identity-initialized linear prediction head with trainable off-diagonal entries | SimSiam-style representation learning · CIFAR-10, CIFAR-100, and STL-10 | linear-classifier accuracy | No clear effect |
| Replace the identity-initialized linear prediction head with a two-layer nonlinear MLP (ablation) | SimSiam-style representation learning · CIFAR-10, CIFAR-100, and STL-10 | linear-classifier accuracy | No clear effect |
| Replace the identity-initialized linear prediction head with a randomly initialized trainable linear head (ablation) | SimSiam-style representation learning · CIFAR-10, CIFAR-100, and STL-10 | linear-classifier accuracy | No clear effect |
| Remove output normalization while training the prediction head (ablation) | Non-contrastive self-supervised learning · Two-feature theoretical setting | learning weak features | Hurt |
| Remove stop-gradient while training the prediction head (ablation) | Non-contrastive self-supervised learning · Two-feature theoretical setting | learning weak features | Hurt |
| Use output normalization and stop-gradient together with the trainable prediction head | Non-contrastive self-supervised learning · Two-feature theoretical setting | learning weak features | Helped |
| Train an identity-initialized prediction head rather than leave it fixed at identity (ablation) | SimSiam-style representation learning · Real-world experiments described in the paper | representation quality | Helped |
Closest papers by meaning
- Understanding self-supervised learning dynamics without contrastive pairs (ICML 2021 · similarity 0.89)
- Dual Perspectives on Non-Contrastive Self-Supervised Learning (ICLR 2026 · similarity 0.88)
- Collapse-Proof Non-Contrastive Self-Supervised Learning (ICML 2025 · similarity 0.86)
- PEIRA: Learning Predictive Encoders through Inter-View Regressor Alignment (NeurIPS 2026 · similarity 0.86)
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data (NeurIPS 2025 · similarity 0.85)
- Contrastive-Equivariant Self-Supervised Learning Improves Alignment with Primate Visual Area IT (NeurIPS 2024 · similarity 0.85)
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.