Research map / Architectures and efficiency / Pruning and efficient tuning
Superposition of many models into one
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 |
|---|---|---|---|
| Use binary parameter superposition instead of a standard network | Fully connected network · Permuting MNIST | accuracy on the initial task | Helped |
| Increase hidden-layer size in a binary-superposition network (ablation) | pspBinary · Permuting MNIST | accuracy on the initial task | Helped |
| Use rotational superposition instead of binary superposition (ablation) | Parameter Superposition · Permuting MNIST | accuracy on the initial task | Helped |
| Use complex superposition instead of binary superposition (ablation) | Parameter Superposition · Permuting MNIST | accuracy on the initial task | Helped |
| Use binary parameter superposition instead of a standard network | Fully connected network · Permuting MNIST, average accuracy on the last ten tasks | accuracy | Helped |
| Use parameter superposition instead of a standard network | Neural network · Rotating FashionMNIST | test accuracy at 0° rotation | Helped |
| Change contexts at every time step and reuse them in the next cycle, rather than use more detailed task identity (ablation) | pspComplex · Rotating FashionMNIST | test accuracy at 0° rotation | Hurt |
| Mix phases of adjacent timepoint contexts (ablation) | pspFast · Rotating FashionMNIST | test accuracy at 0° rotation | Helped |
| Use binary parameter superposition with one output layer instead of a standard ResNet-18 | ResNet-18 · iCIFAR, CIFAR-10 test set | accuracy | Helped |
| Use a separate output head for each task instead of a single output head (ablation) | ResNet-18 · iCIFAR, CIFAR-10 test set | accuracy | Helped |
Closest papers by meaning
- Model Preserving Compression for Neural Networks (NeurIPS 2022 · similarity 0.85)
- In defense of parameter sharing for model-compression (ICLR 2024 · similarity 0.85)
- Forget the Data and Fine-Tuning! Just Fold the Network to Compress (ICLR 2025 · similarity 0.84)
- Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models (NeurIPS 2024 · similarity 0.83)
- Strong Model Collapse (ICLR 2025 · similarity 0.83)
- Superposition Yields Robust Neural Scaling (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.