Research map / Architectures and efficiency / RNNs and normalization
How to Train your HIPPO: State Space Models with Generalized Orthogonal Basis Projections
Albert Gu, Isys Johnson, Aman Timalsina, Atri Rudra, Christopher Re
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 |
|---|---|---|---|
| Use the LegS initialization instead of original S4 | S4 · LRA average | accuracy: 80.48 -> 86.09 | Improved |
| Use the LegS initialization instead of original S4 | S4 · Path-X | accuracy: 88.10 -> 96.35 | Improved |
| Use the FouT initialization instead of LegS | S4 · Image | accuracy: 88.65 -> 89.07 | Improved |
| Use the FouT initialization instead of LegS | S4 · Pathfinder | accuracy: 94.20 -> 94.46 | Improved |
| Use the FouT initialization instead of LegS | S4 · Path-X | accuracy | Hurt |
| Initialize half the kernels as LegS and half as FouT rather than using FouT alone | S4-FouT · LRA average | accuracy: 77.90 -> 78.50 | Improved |
| Set the timescale Δ to match known task dependencies | S4 variants · Path-X | accuracy | Improved |
| Set Δ to the theoretically optimal value for the 1000-step lag | S4-FouT · Delay Task | RMSE | Improved |
| Set Δ incorrectly rather than matching the task timescale | S4-FouT · Delay Task | RMSE | Hurt |
| Initialize C with standard deviation 1.0 rather than 0.01 (ablation) | S4-LegS · Sequential CIFAR | accuracy: 85.91 (0.41) -> 86.49 (0.51) | Improved |
| Initialize C with standard deviation 1.0 rather than 0.1 (ablation) | S4-LegS · Sequential CIFAR | accuracy: 86.33 (0.01) -> 86.49 (0.51) | No clear effect |
| Initialize C with standard deviation 1.0 rather than 10.0 (ablation) | S4-LegS · Sequential CIFAR | accuracy: 84.40 (0.16) -> 86.49 (0.51) | Improved |
Closest papers by meaning
- Autocorrelation Matters: Understanding the Role of Initialization Schemes for State Space Models (ICLR 2025 · similarity 0.87)
- From Generalization Analysis to Optimization Designs for State Space Models (ICML 2024 · similarity 0.87)
- Uncovering the Spectral Bias in Diagonal State Space Models (NeurIPS 2025 · similarity 0.86)
- UnHiPPO: Uncertainty-aware Initialization for State Space Models (ICML 2025 · similarity 0.86)
- HOPE for a Robust Parameterization of Long-memory State Space Models (ICLR 2025 · similarity 0.86)
- Structured Sparse Transition Matrices to Enable State Tracking in State-Space Models (NeurIPS 2025 · similarity 0.86)
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.
