Research map / Probabilistic and scientific modeling / Dynamical systems
Beyond Geometry: Comparing the Temporal Structure of Computation in Neural Circuits with Dynamical Similarity Analysis
Mitchell Ostrow, Adam Joseph Eisen, Leo Kozachkov, Ila R Fiete
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 |
|---|---|---|---|
| remove delay embedding and reduced-rank regression (use standard or delay-embedded Procrustes, or DSA without delay embedding) (ablation) | DSA · Sine Wave task, RNN vs GRU | similarity of RNN and GRU | Hurt |
| compare purely temporal information via power spectrum Procrustes instead of DSA (ablation) | DSA · Sine Wave task, RNN vs GRU | angular distance: 1.41 | Hurt |
| use DSA instead of Procrustes to compare FlipFlop RNNs by architecture and activation | Procrustes shape analysis · 3-bit FlipFlop, 240 RNNs | linear SVM accuracy on architecture/activation labels | Improved |
| use DSA instead of Procrustes to distinguish three network dynamics with identical condition-averaged trajectories | Procrustes shape analysis · Decision-making networks (bistable, line, fixed point) | linear classifier accuracy | Improved |
| deform ring attractor geometry by increasing beta from 0.1 to 4 | DSA vs Procrustes · Ring attractor | distance to beta=0.1: 0.05 to 0.12 | Improved |
| convert ring attractor to line attractor by ablating neurons (alpha below 1) | DSA · Ring attractor | distance to alpha=0 | Improved |
| use DSA on trajectories of 21 aggregate statistics instead of the prior linear SVM | prior learning-rule classifier · Learning rule classification (Adam, SGD momentum, IA, FA) | test accuracy (%): 55 -> 88.5 | Improved |
| vary DSA hyperparameters widely (ablation) | DSA · Learning rule classification (CNNs) | test accuracy (%): 88.5 -> 81 | Mixed |
| vary DSA hyperparameters widely (ablation) | DSA · 3-layer linear nets learning rules | test accuracy (%): 89 -> 80 | Mixed |
| increase HAVOK rank and number of delays (ablation) | HAVOK model in DSA · FlipFlop HAVOK grid sweep | R2, MSE, Pearson correlation | Improved |
Closest papers by meaning
- RNNs perform task computations by dynamically warping neural representations (NeurIPS 2025 · similarity 0.91)
- Geometry-Aware Similarity Metrics for Neural Representations on Riemannian and Statistical Manifolds (NeurIPS 2026 · similarity 0.91)
- Comparing noisy neural population dynamics using optimal transport distances (ICLR 2025 · similarity 0.89)
- Beyond DSA: Conjugacy-based Comparison of Dynamical Systems (NeurIPS 2026 · similarity 0.89)
- Identifying Equivalent Training Dynamics (NeurIPS 2024 · similarity 0.88)
- Inference of Neural Dynamics Using Switching Recurrent Neural Networks (NeurIPS 2024 · similarity 0.88)
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.
