Research map / Probabilistic and scientific modeling / Dynamical systems
Neural Metamorphosis
Xingyi Yang, Xinchao Wang
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 |
|---|---|---|---|
| replace layer-wise TSP permutation (Greedy) with mSHP weight permutation (ablation) | NeuMeta · CIFAR10 | accuracy: 91.67 -> 91.87 | Improved |
| replace layer-wise TSP 2-Opt permutation with mSHP (ablation) | NeuMeta · CIFAR10 | accuracy: 91.31 -> 91.87 | Improved |
| add mSHP weight permutation versus no permutation (ablation) | NeuMeta · CIFAR10 | accuracy: 91.62 -> 91.87 | Improved |
| add mSHP weight permutation versus no permutation (ablation) | NeuMeta · CIFAR10 | total variation: 58141.80 -> 45576.02 | Improved |
| add residual connections to the INR (ablation) | NeuMeta INR · CIFAR10 | accuracy: 90.18 -> 91.87 | Improved |
| use block-based INR with separate MLPs per parameter (ablation) | NeuMeta INR · CIFAR10 | accuracy: 91.74 -> 91.87 | Improved |
| add EMA with gamma 0.995 during INR training (ablation) | NeuMeta INR · CIFAR10 | accuracy: 91.32 -> 91.87 | Improved |
| increase EMA gamma to 0.999 from 0.995 (ablation) | NeuMeta INR · CIFAR10 | accuracy: 91.87 -> 91.42 | Hurt |
| increase INR depth to 12 layers with 256 hidden units (ablation) | NeuMeta INR · CIFAR10 | accuracy: 91.87 -> 91.27 | Hurt |
| use 8 layers, 512 hidden units, 32 frequencies versus 5 layers, 256 hidden, 32 frequencies (ablation) | NeuMeta INR · CIFAR10 | accuracy: 91.23 -> 91.87 | Improved |
| extend morphing to network depth, evaluating unseen depths beyond training range 1-3 | NeuMeta · MNIST depth morphing | accuracy: about 99% in-range; 56.97% at L1=L2=7 | Mixed |
Closest papers by meaning
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.78)
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.77)
- Inpainting the Neural Picture: Inferring Unrecorded Brain Area Dynamics from Multi-Animal Datasets (NeurIPS 2025 · similarity 0.77)
- Meta-Dynamical State Space Models for Integrative Neural Data Analysis (ICLR 2025 · similarity 0.77)
- TRENDy: Temporal Regression of Effective Nonlinear Dynamics (ICLR 2025 · similarity 0.77)
- Convergent Functions, Divergent Forms (NeurIPS 2025 · similarity 0.76)
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.
