Research map / Probabilistic and scientific modeling / Dynamical systems
Learning system dynamics without forgetting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Formulates continual dynamics learning (CDL) for graph-ODE trajectory prediction and proposes MS-GODE, a sub-network (edge-popup mask) model with reconstruction-based mode switching, plus a new biological benchmark Bio-CDL built from VCell.
Problem
Continual learning of trajectory prediction models across systems with evolving dynamics (different interaction types and strengths, different object counts), where catastrophic forgetting occurs and existing CL methods fail under irregular, incomplete observations.
Approach
Freeze randomly initialized backbone weights (encoder, ODE generator, decoder built on LG-ODE) and learn a binary edge-popup mask per system, then select the mask at test time by reconstructing the second half of an observation from the first half, instead of entropy-based selection used by SupSup.
Key results
- On physics Seq3 (high-level dynamics shift), MS-GODE reaches AP 0.113±0.001 with AF -0.000, versus joint training at 0.116±0.009 and EWC at 0.148±0.011.
- On Bio-CDL EGFR1→Ran1→EGFR2→Ran2, MS-GODE achieves AP 0.144±0.012 and AF -0.003±0.036, outperforming BCMR (0.149, AF 0.013) and SDP (0.197); joint training reaches 0.055.
- On low-level shift Seq1, MS-GODE AP 0.200±0.003 vs. EWC 0.208±0.015 and joint 0.194±0.006, with AF 0.002.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MS-GODE instead of joint training | LG-ODE on Physics Sequence 1 · Physics Sequence 1 | AP: 0.194$\pm$0.006 -> 0.200$\pm$0.003 | Hurt |
| Use MS-GODE instead of joint training | LG-ODE on Physics Sequence 2 · Physics Sequence 2 | AP: 0.186$\pm$0.015 -> 0.204$\pm$0.005 | Hurt |
| Use MS-GODE instead of joint training | LG-ODE on Physics Sequence 3 · Physics Sequence 3 | AP: 0.116$\pm$0.009 -> 0.113$\pm$0.001 | Helped |
| Use MS-GODE instead of joint training | LG-ODE on Bio-CDL cellular sequence · Bio-CDL cellular sequence | AP: 0.055$\pm$<0.001 -> 0.144$\pm$0.012 | Hurt |
| Use fast selection rather than top-k selection for mask binarization (ablation) | MS-GODE · Bio-CDL and physics system sequences | prediction error | Helped |
| Decrease dropout rate while using mask selection (ablation) | MS-GODE · system sequences | prediction error | Helped |
Surprising findings
- Top-k mask binarization was inferior to fast (threshold-free) selection, and was more sensitive on cellular systems than physics systems.
- Gradual dynamics shift sequences (Seq1, Seq2) were harder to learn than abrupt-shift sequences, as indicated by higher AP.
Limitations
- Bio-CDL data is not yet released and must be regenerated through the VCell platform manually.
- Mode selection errors remain a source of performance drop on gradual shift sequences.
- Pre-trained backbones were not used because suitable pre-trained dynamics models were unavailable for extrapolation.
Experimental setup
Datasets: spring-connected particle simulations charged particle simulations Bio-CDL (VCell EGFR and Ran systems)
Benchmarks: Physics system sequences (Seq1-3) Bio-CDL EGFR-Ran sequence
Metrics: MSE average performance (AP) average forgetting (AF)
Baselines: Fine-tune EWC GEM LwF BCMR SDP Joint training
Builds on: LG-ODE edge-popup algorithm SupSup Neural Relational Inference Elastic Weight Consolidation
Closest papers by meaning
- A Regime-Aware Trajectory Prediction Framework for 1000+ Systems Biology Models (ICML 2026 · similarity 0.88)
- Tree-Structured Recurrent Switching Linear Dynamical Systems for Multi-Scale Modeling (ICLR 2019 · similarity 0.88)
- Skip the Equations: Learning Behavior of Personalized Dynamical Systems Directly From Data (ICML 2025 · similarity 0.86)
- Learning Interpretable Hierarchical Dynamical Systems Models from Time Series Data (ICLR 2025 · similarity 0.85)
- Learning Hybrid Dynamics Models with Simulator-Informed Latent States (AAAI 2024 · similarity 0.85)
- Neural Persistence Dynamics (NeurIPS 2024 · 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.