Research map / Architectures and efficiency / RNNs and normalization
System Identification with Time-Aware Neural Sequence Models
Thomas Demeester
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 |
|---|---|---|---|
| Add normalized step size δ_n as an input | Standard GRU · CSTR (missing data) | RRSE (%): 10.8±0.6 -> 9.2±1.4 | Improved |
| Add normalized step size δ_n as an input | Standard GRU · Winding (missing data) | RRSE (%): 30.2 ± 0.4 -> 28.9 ± 1.7 | Improved |
| Use an incremental update without unit-root compensation instead of the stationary update (ablation) | Time-aware GRU · CSTR (missing data) | RRSE (%): 12.1±1.3 -> 79.7±8.0 | Hurt |
| Use an incremental update without unit-root compensation instead of the stationary update (ablation) | Time-aware GRU · Winding (missing data) | RRSE (%): 33.1±0.6 -> 64.8 ± 10.2 | Hurt |
| Increase Runge-Kutta order from Euler to RK4 (ablation) | Stationary time-aware GRU with constant input interpolation · CSTR (missing data) | RRSE (%): 12.1±1.3 -> 8.0±0.4 | Improved |
| Use linear rather than constant input interpolation (ablation) | Stationary time-aware GRU with Midpoint scheme · Winding (missing data) | RRSE (%): 35.3±2.1 -> 28.2±1.8 | Improved |
| Use linear rather than constant input interpolation (ablation) | Stationary time-aware GRU with Kutta3 scheme · Winding (missing data) | RRSE (%): 34.7±4.3 -> 27.1±1.0 | Improved |
| Use linear rather than constant input interpolation (ablation) | Stationary time-aware GRU with RK4 scheme · Winding (missing data) | RRSE (%): 32.2±6.6 -> 25.6±1.4 | Improved |
| Increase Runge-Kutta order from Euler to RK4 while using constant input interpolation (ablation) | Stationary time-aware GRU · Winding (missing data) | RRSE (%): 33.1±0.6 -> 32.2±6.6 | No clear effect |
| Use a time-aware Euler update rather than a standard GRU that ignores missing times | GRU · CSTR (missing data) | RRSE (%): 10.8±0.6 -> 12.1±1.3 | Hurt |
| Use a time-aware Euler update rather than a standard GRU that ignores missing times | GRU · Winding (missing data) | RRSE (%): 30.2 ± 0.4 -> 33.1±0.6 | Hurt |
Closest papers by meaning
- Transformation of ReLU-based recurrent neural networks from discrete-time to continuous-time (ICML 2020 · similarity 0.88)
- Identifying nonlinear dynamical systems with multiple time scales and long-range dependencies (ICLR 2021 · similarity 0.86)
- Stable Recurrent Models (ICLR 2019 · similarity 0.86)
- Parallel-in-Time Training of Recurrent Neural Networks for Dynamical Systems Reconstruction (NeurIPS 2026 · similarity 0.85)
- Learning dynamics in linear recurrent neural networks (ICML 2025 · similarity 0.84)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.84)
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.
