Research map / Time series / Causal and generative time series
HiPPO: Recurrent Memory with Optimal Polynomial Projections
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 HiPPO-LegS polynomial-projection memory instead of recurrent baselines | pMNIST sequence classification · pMNIST validation | accuracy (%): 98.34 vs 97.08 (LMU) | Helped |
| Set LegT timescale θ to 20 rather than the sequence-length value (ablation) | HiPPO-LegT · pMNIST validation | accuracy (%): 98.00 -> 91.75 | Hurt |
| Set LegT timescale θ to 2000 rather than the sequence-length value (ablation) | HiPPO-LegT · pMNIST validation | accuracy (%): 98.00 -> 97.90 | Hurt |
| Set LagT step size Δt to 0.1 instead of 1.0 (ablation) | HiPPO-LagT · pMNIST validation | accuracy (%): 98.15 -> 96.44 | Hurt |
| Set LagT step size Δt to 0.01 instead of 1.0 (ablation) | HiPPO-LagT · pMNIST validation | accuracy (%): 98.15 -> 90.71 | Hurt |
| Replace framework-derived memory dynamics with random A and B matrices (ablation) | HiPPO recurrent memory · pMNIST validation | accuracy (%): 98.34 -> 69.93 | Hurt |
| Use HiPPO-LegS rather than recurrent and sequence-model baselines | Pixel-by-pixel image classification · Pixel-by-pixel image classification test | accuracy (%): 98.3 vs 98.13 (TrellisNet); 98.3 vs 97.58 (URLSTM + Zoneout) | Mixed |
| Use random dynamics instead of the HiPPO-derived dynamics (ablation) | HiPPO memory · IMDB test | accuracy (%): 87.8 $\pm$ 0.2 -> 82.9 $\pm$ 0.3 | Hurt |
| Vary LegT timescale θ across 100, 1000, and 10000 (ablation) | HiPPO-LegT · IMDB test | accuracy (%): 87.4 $\pm$ 0.3; 87.7 $\pm$ 0.2; 87.9 $\pm$ 0.3 | No clear effect |
| Use HiPPO-LegS rather than LSTM on IMDB | IMDB sentiment classification · IMDB test | accuracy (%): 87.8 $\pm$ 0.2 vs 87.3 $\pm$ 0.4 | No clear effect |
Closest papers by meaning
- HiPPO Zoo: Explicit Memory Mechanisms for Interpretable State Space Models (ICML 2026 · similarity 0.91)
- Recurrent Memory for Online Interdomain Gaussian Processes (NeurIPS 2025 · similarity 0.89)
- Practical Real Time Recurrent Learning with a Sparse Approximation (ICLR 2021 · similarity 0.87)
- Memory-Gated Recurrent Networks (AAAI 2021 · similarity 0.86)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.86)
- Exploring the Promise and Limits of Real-Time Recurrent Learning (ICLR 2024 · 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.