Research map / Video / Video segmentation and tracking
LayerLock: Non-collapsing Representation Learning with Progressive Freezing
Goker Erdogan, Nikhil Parthasarathy, Catalin Ionescu, Drew A. Hudson, Alexander Lerchner, Andrew Zisserman, Mehdi S. M. Sajjadi, Joao Carreira
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 |
|---|---|---|---|
| add LayerLock progressive freezing | 4DS-G MAE · SSv2 | top-1 accuracy: 63.1 -> 66.1 | Improved |
| add LayerLock progressive freezing | 4DS-e MAE · SSv2 | top-1 accuracy: 64.6 -> 67.1 | Improved |
| add LayerLock progressive freezing | 4DS-e MAE · ScanNet depth | AbsRel (x10): 0.94 -> 0.98 | Hurt |
| add LayerLock to V-JEPA-L | V-JEPA-L · SSv2 | top-1 accuracy: 52.1 -> 57.0 | Improved |
| add LayerLock to V-JEPA-L | V-JEPA-L · Kinetics-700-2020 | top-1 accuracy: 42.5 -> 43.5 | Improved |
| progressive freezing alone on MAE (efficiency) (ablation) | 4DS-G MAE ViT-G 250M · SSv2 | top-1 accuracy: 56.1 -> 56.0 | No clear effect |
| compute latent loss on only 5% of patches instead of 100% (ablation) | 4DS-G LayerLock · SSv2 | top-1 accuracy: 66.1 -> 64.9 | Hurt |
| compute latent loss on only 5% of patches instead of 100% (ablation) | 4DS-G LayerLock · ScanNet depth | AbsRel (x10): 1.00 -> 1.12 | Hurt |
| add latent loss to MAE without freezing (constant weight) (ablation) | 4DS-H MAE · SSv2 | top-1 accuracy: 50.1 -> 3.7 | Hurt |
| add latent loss to MAE without freezing (cosine weight) (ablation) | 4DS-H MAE · SSv2 | top-1 accuracy: 50.1 -> 5.6 | Hurt |
| add 3D RoPE (ablation) | 4DS-G · SSv2 | top-1 accuracy: 56.1 -> 58.9 | Improved |
| keep multiple targets instead of single latest (ablation) | LayerLock · SSv2 | top-1 accuracy: 55.5 -> 55.4 | No clear effect |
Closest papers by meaning
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.86)
- EVEREST: Efficient Masked Video Autoencoder by Removing Redundant Spatiotemporal Tokens (ICML 2024 · similarity 0.86)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.85)
- Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks (NeurIPS 2026 · similarity 0.85)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.85)
- VideoMAC: Video Masked Autoencoders Meet ConvNets (CVPR 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.
