Research map / Continual learning / Online and replay-based
Supermasks in Superposition
Mitchell Wortsman, Vivek Ramanujan, Rosanne Liu, Aniruddha Kembhavi, Mohammad Rastegari, Jason Yosinski, Ali Farhadi
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 |
|---|---|---|---|
| Initialize each new mask from the running mean of previous task masks (ablation) | SupSup on SplitCIFAR100 · SplitCIFAR100 | Avg Acc@1: 89.57~\pm 0.64 -> \textbf{91.66} $~\pm 0.74$ | Improved |
| Use a running mean initialization for subsequent task masks (ablation) | SupSup on SplitCIFAR · SplitCIFAR100 | training iterations for comparable accuracy | Improved |
| Use output size s=200 rather than s=100 (ablation) | SupSup with LeNet 300-100 · PermutedMNIST (250 tasks, Avg) | accuracy: 93.66 -> 93.47 | No clear effect |
| Use output size s=25 rather than s=100 (ablation) | SupSup with LeNet 300-100 · PermutedMNIST (250 tasks, Avg) | accuracy: 93.66 -> 93.27 | Hurt |
| Use output size s=200 rather than s=100 (ablation) | SupSup with FC 1024-1024 · PermutedMNIST (250 tasks, Avg) | accuracy: 95.50 -> 96.07 | Improved |
| Use output size s=25 rather than s=100 (ablation) | SupSup with FC 1024-1024 · PermutedMNIST (250 tasks, Avg) | accuracy: 94.04 -> 65.34 | Hurt |
| Use the G objective instead of entropy H for task inference (ablation) | SupSup with LeNet 300-100 · PermutedMNIST (250 tasks, Avg) | accuracy: 93.66 -> 93.66 | No clear effect |
| Use the G objective instead of entropy H for task inference (ablation) | SupSup with FC 1024-1024 · PermutedMNIST (250 tasks, Avg) | accuracy: 95.98 -> 96.07 | Improved |
| Use one-shot rather than binary task inference (ablation) | SupSup · RotatedMNIST | accuracy | Hurt |
| Infer task identity without any task identity during training or inference (ablation) | SupSup · PermutedMNIST (2500 tasks) | accuracy: 93.45 -> 92.54 | Hurt |
| Use entropy H rather than the confidence alternative M for ABatchE inference (ablation) | ABatchE · PermutedMNIST | accuracy | Improved |
| Use a larger batch for ABatchE task inference (ablation) | ABatchE · PermutedMNIST | forgetting: 16 images | Improved |
Closest papers by meaning
- Signing the Supermask: Keep, Hide, Invert (ICLR 2022 · similarity 0.87)
- Multicoated Supermasks Enhance Hidden Networks (ICML 2022 · similarity 0.84)
- Towards Exact Gradient-based Training on Analog In-memory Computing (NeurIPS 2024 · similarity 0.83)
- Boosting Order-Preserving and Transferability for Neural Architecture Search: a Joint Architecture Refined Search and Fine-tuning Approach (CVPR 2024 · similarity 0.83)
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (NeurIPS 2026 · similarity 0.82)
- Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks (ICML 2025 · similarity 0.82)
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.
