Research map / Label-efficient and robust learning / Recognition and long-tail
Split-Brain Autoencoders: Unsupervised Learning by Cross-Channel Prediction
Check 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 |
|---|---|---|---|
| replace full-input reconstruction with two disjoint cross-channel prediction branches (ablation) | AlexNet unsupervised pretraining · ImageNet classification, conv5 | accuracy: 11.5 -> 32.3 | Helped |
| add random 50% input dropout to full-input autoencoder pretraining (ablation) | Lab→Lab · ImageNet classification, conv5 | accuracy: 11.5 -> 12.3 | Helped |
| use classification rather than regression loss in both split branches (ablation) | Split-Brain Auto · ImageNet classification, conv5 | accuracy: 32.3 -> 32.8 | Helped |
| use classification rather than regression loss for colorization (ablation) | L→ab colorization encoder · ImageNet classification, conv5 | accuracy: 30.1 -> 32.0 | Helped |
| use regression rather than classification loss for grayscale prediction (ablation) | ab→L grayscale prediction encoder · ImageNet classification, conv5 | accuracy: 19.2 -> 21.7 | Helped |
| add a complementary grayscale-from-color branch to colorization (ablation) | Split-Brain Auto · ImageNet classification, conv5 | accuracy: 32.0 -> 32.8 | Helped |
| use two disjoint concatenated branches instead of a shared network for both prediction directions (ablation) | cross-channel encoders · ImageNet classification, conv5 | accuracy: 21.1 -> 32.3 | Helped |
| add a full-input autoencoder objective to the shared bidirectional prediction network (ablation) | (L,ab)→(ab,L) · ImageNet classification, conv5 | accuracy: 21.1 -> 18.9 | Hurt |
| replace opposite-direction prediction with two colorization branches (ablation) | Split-Brain Auto · ImageNet classification, conv5 | accuracy: 31.3 -> 32.8 | Hurt |
| use split-brain RGB-D cross-modal pretraining | NYU-D detection representations · NYU-D detection, RGB-D ensemble | performance: 34.5 -> 38.1 | Helped |
Closest papers by meaning
- Supervised autoencoders: Improving generalization performance with unsupervised regularizers (NeurIPS 2018 · similarity 0.85)
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.85)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.83)
- A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders (NeurIPS 2025 · similarity 0.83)
- Mechanistic Interpretability with Sparse Autoencoder Neural Operators (NeurIPS 2026 · similarity 0.83)
- Auto-Compressing Networks (NeurIPS 2025 · 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.