Research map / Probabilistic and scientific modeling / Dynamical systems
The geometry of integration in text classification RNNs
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 categorical labels with N mutually exclusive classes | Text-classification RNNs · Categorical classification | integration manifold geometry and dimensionality | Helped |
| Use ordered class labels | Text-classification RNNs · 3-class and 5-class ordered classification | integration manifold dimensionality | Mixed |
| Track independent binary decisions instead of one exclusive class | Text-classification RNNs · Multi-label classification | integration manifold geometry and dimensionality | Helped |
| Use sentiment-only labels rather than labels based on sentiment and intensity (ablation) | Synthetic ordered-classification RNN · Ordered classification with N>2 | match to natural-data dynamics | Hurt |
| Increase ℓ2 regularization (ablation) | Categorical synthetic RNNs · Categorical synthetic data | hidden-state dimensionality | Hurt |
| Decrease ℓ2 regularization (ablation) | Categorical synthetic RNNs · Categorical synthetic data | hidden-state dimensionality | Mixed |
| Balance the number of examples per class (ablation) | GoEmotions RNN · 2-label GoEmotions | match between synthetic and natural dynamics | Helped |
| Draw synthetic phrases uniformly over words rather than uniformly over possible scores (ablation) | Synthetic categorical RNN · Synthetic categorical data | qualitative dynamics | No clear effect |
Closest papers by meaning
- Reverse engineering recurrent networks for sentiment classification reveals line attractor dynamics (NeurIPS 2019 · similarity 0.91)
- RNNs perform task computations by dynamically warping neural representations (NeurIPS 2025 · similarity 0.87)
- Learning dynamics in linear recurrent neural networks (ICML 2025 · similarity 0.87)
- The Simplicity Bias in Multi-Task RNNs: Shared Attractors, Reuse of Dynamics, and Geometric Representation (NeurIPS 2023 · similarity 0.86)
- Detecting Invariant Manifolds in ReLU-Based RNNs (ICLR 2026 · similarity 0.86)
- Correlation Dimension of Autoregressive Large Language Models (NeurIPS 2025 · 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.