Research map / Learning theory and algorithms / Sample complexity
Near-optimal Offline and Streaming Algorithms for Learning Non-Linear Dynamical Systems
Suhas S Kowshik, Dheeraj Mysore Nagaraj, Prateek Jain, Praneeth Netrapalli
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 |
|---|---|---|---|
| Use Quasi Newton updates normalized by the empirical second-moment matrix instead of mixing-time-based subsampling | Offline NLDS parameter estimation · Non-mixing NLDS | squared Frobenius parameter error: \frac{d^2\tmix}{T} -> \frac{d^2\sigma^2}{T\lmin{\hat{G}}} | Improved |
| Use Quasi Newton updates for offline estimation under finite fourth-moment noise | Offline NLDS parameter estimation · Mixing NLDS | squared Frobenius parameter error: \frac{d^2\tmix}{T} -> \frac{d^2\sigma^2}{T\lmin{G}} | Improved |
| Process samples in reverse order within buffers, with gaps between buffers | SGD-RER · Synthetic leaky-ReLU NLDS | Frobenius parameter error | Improved |
| Use SGD-RER instead of forward SGD or random experience replay | Streaming NLDS estimation · Synthetic leaky-ReLU NLDS | Frobenius parameter error | Improved |
| Increase the SGD-RER buffer size beyond the tested robust range (ablation) | SGD-RER · Synthetic NLDS, buffer-size experiment | Frobenius parameter error | Hurt |
| Use a non-expansive ReLU link for parameter recovery | NLDS parameter estimation · ReLU NLDS | minimax loss: c_0\min\left(1,\frac{\exp(c_1d)}{T}\right) | Hurt |
| Use GLMtron rather than Quasi Newton updates | Offline NLDS estimation · Synthetic leaky-ReLU NLDS | convergence time | Hurt |
| Use SGD-RER with heavy-tailed Student t noise | SGD-RER · Synthetic NLDS with Student t noise, scale ν = 4.1 | Frobenius parameter error | No clear effect |
Closest papers by meaning
- Learning with little mixing (NeurIPS 2022 · similarity 0.87)
- Streaming Linear System Identification with Reverse Experience Replay (NeurIPS 2021 · similarity 0.87)
- Sparse Learning of Dynamical Systems in RKHS: An Operator-Theoretic Approach (ICML 2023 · similarity 0.87)
- A Memory Efficient Unified Algorithm for Online Learning of Linear Dynamical Systems (NeurIPS 2026 · similarity 0.85)
- Sharp Rates in Dependent Learning Theory: Avoiding Sample Size Deflation for the Square Loss (ICML 2024 · similarity 0.85)
- Parallel-in-Time Training of Recurrent Neural Networks for Dynamical Systems Reconstruction (NeurIPS 2026 · similarity 0.85)
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.
