Research map / Probabilistic and scientific modeling / Molecules and proteins
Twin Networks: Matching the Future for Sequence Generation
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 |
|---|---|---|---|
| Add TwinNet with learned affine mapping | Baseline speech recognition model · WSJ test | CER (%): 6.8 -> 6.2 | Helped |
| Use identity mapping instead of learned g (ablation) | TwinNet speech recognition model · WSJ test | CER (%): 6.2 -> 6.6 | Hurt |
| Train forward states to predict Gaussian noise rather than backward states (ablation) | Baseline speech recognition model · WSJ test | CER (%): 6.8 -> 6.9 | Hurt |
| Set backward states to zero (activation regularization) (ablation) | Baseline speech recognition model · WSJ test | CER (%): 6.8 -> 6.5 | Helped |
| Add TwinNet | Show & Tell with ResNet-101 · MS COCO, no attention | CIDEr: 84.3 -> 87.7 | Helped |
| Add TwinNet | Soft Attention with ResNet-101 · MS COCO, attention | CIDEr: 89.2 -> 96.2 | Helped |
| Add TwinNet | Show & Tell with ResNet-152 · MS COCO, no attention | CIDEr: 93.0 -> 94.7 | Helped |
| Add TwinNet | Soft Attention with ResNet-152 · MS COCO, attention | METEOR: 25.3 -> 25.2 | Hurt |
| Add TwinNet regularization | 3-layer LSTM · Sequential MNIST test | negative log-likelihood (nats): 79.87 -> 79.35 | Helped |
| Add dropout (ablation) | 3-layer LSTM · Sequential MNIST test | negative log-likelihood (nats): 79.87 -> 79.59 | Helped |
| Add TwinNet to the dropout baseline | 3-layer LSTM with dropout · Sequential MNIST test | negative log-likelihood (nats): 79.59 -> 79.12 | Helped |
| Add TwinNet regularization | AWD-LSTM · Penn Treebank test | perplexity: 58.8 -> 58.3 | Helped |
Closest papers by meaning
- Learning and Generalization in RNNs (NeurIPS 2021 · similarity 0.87)
- Multiresolution Recurrent Neural Networks: An Application to Dialogue Response Generation (AAAI 2017 · similarity 0.87)
- Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks (ICML 2018 · similarity 0.86)
- Prospective Coding Improves Learning in Deep Continuous-Time Recurrent Networks (NeurIPS 2026 · similarity 0.85)
- RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling (NeurIPS 2025 · similarity 0.85)
- Predictive Attractor Models (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.