Research map / Label-efficient and robust learning / Self-supervised and contrastive
Evolving Losses for Unsupervised Video Representation Learning
AJ Piergiovanni, Anelia Angelova, Michael S. Ryoo
Check 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 |
|---|---|---|---|
| Evolve the combination and weights of self-supervised and distillation losses using unsupervised Zipf matching rather than a randomly sampled loss (ablation) | ELo · HMDB51 | fine-tune accuracy: 31.2 -> 67.4 | Improved |
| Add distillation of other modalities into the RGB network (ablation) | ELo · HMDB51 | fine-tune accuracy: 53.7 -> 67.4 | Improved |
| Add distillation of other modalities into the RGB network (ablation) | ELo · UCF101 | fine-tune accuracy: 84.2 -> 93.8 | Improved |
| Use label-guided HMDB clustering fitness instead of unsupervised Zipf-matching fitness (ablation) | ELo · HMDB51 | fine-tune accuracy: 67.4 -> 67.8 | No clear effect |
| Use CMA-ES instead of tournament selection for loss-weight search (ablation) | Loss evolution · HMDB | best loss fitness accuracy: 61.4 -> 67.4 | Improved |
| Use CMA-ES for 250 iterations rather than random search for 2000 iterations (ablation) | Loss evolution · HMDB | best loss fitness accuracy: 52.4 -> 67.4 | Improved |
| Increase the amount of unlabeled video data while holding training iterations fixed | ELo · HMDB | accuracy | Improved |
| Increase unlabeled data and training time by holding the number of epochs fixed | ELo · HMDB | accuracy | Improved |
| Use unsupervised ELo pretraining before Kinetics fine-tuning instead of random initialization | ELo · Kinetics-400 | accuracy: 71.2 -> 74.4 | Improved |
| Use unsupervised ELo pretraining before Kinetics fine-tuning instead of Frame Shuffle pretraining | ELo · Kinetics-400 | accuracy: 70.9 -> 74.4 | Improved |
Closest papers by meaning
- Unsupervised Learning From Video With Deep Neural Embeddings (CVPR 2020 · similarity 0.87)
- Unsupervised Learning of View-invariant Action Representations (NeurIPS 2018 · similarity 0.87)
- Trading robust representations for sample complexity through self-supervised visual experience (NeurIPS 2018 · similarity 0.87)
- Memory-augmented Dense Predictive Coding for Video Representation Learning (ECCV 2020 · similarity 0.87)
- Unsupervised Learning of Long-Term Motion Dynamics for Videos (CVPR 2017 · similarity 0.86)
- Temporally Consistent Unbalanced Optimal Transport for Unsupervised Action Segmentation (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.
