Research map / Architectures and efficiency / RNNs and normalization
Multi-Task Recurrent Modular Networks
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 dynamically routed modular encoder-decoder connections rather than the strongest listed fixed or alternate MTL baseline | mtl-RMN · tfMRI Group-1, Tasks 1–3 | accuracy: 89.2±0.6 -> 90.8±2.1; 95.7±1.1 -> 96.7±1.6; 91.3±1.7 -> 92.9±2.4 | Helped |
| Use dynamically routed modular encoder-decoder connections rather than RRNs | mtl-RMN · tfMRI Group-2, Tasks 1, 3, and 4 | accuracy: 75.4±2.2 -> 76.1±1.8; 83.2±3.0 -> 84.4±2.4; 82.6±3.8 -> 83.5±2.2 | Helped |
| Train a modular encoder on related tasks, freeze it, and use it for a held-out task | RMNtr with LSTM · tfMRI unseen-task scenarios A–D | accuracy: 82.6±1.0 -> 83.3±2.3; 82.3±0.6 -> 85.6±1.7; 93.7±0.6 -> 97.0±1.8; 71.7±1.4 -> 79.3±2.6 | Helped |
| Use the trained modular encoder instead of an untrained encoder | RMNtr · tfMRI unseen-task scenarios A–D | accuracy: 82.9±1.2 -> 83.3±2.3; 84.5±1.5 -> 85.6±1.7; 95.3±0.8 -> 97.0±1.8; 73.1±1.7 -> 79.3±2.6 | Helped |
| Add MT-RMN to the underlying multi-task recurrent model | Hindi-English POS tagging models · Hindi-English POS tagging | accuracy: 44.3±0.7 -> 55.2±1.2 (NS-MTL); 46.0±1.9 -> 51.7±1.9 (Cross-stitch); 58.7±0.7 -> 61.5±1.7 (Sluice); 62.8±0.6 -> 64.5±2.1 | Helped |
| Add MT-RMN to the underlying multi-task recurrent model | Hindi-English POS tagging models · Hindi-English POS tagging | F1: 38.9±1.6 -> 53.4±2.4 (NS-MTL); 12.7±0.8 -> 21.4±3.2 (Cross-stitch); 55.4±2.1 -> 57.2±1.8 (Sluice); 46.6±0.4 -> 62.6±2.4 | Helped |
| Replace the layered sub-encoder architecture with one recursive set of sub-encoders (ablation) | MT-RMN · Hindi-English POS tagging with NS-MTL, Sluice, and GIRNet | accuracy: 55.2±1.2 -> 53.5±1.9; 61.5±1.7 -> 60.4±1.2; 64.5±2.1 -> 63.6±1.7 | Hurt |
| Replace the layered sub-encoder architecture with one recursive set of sub-encoders (ablation) | MT-RMN · Hindi-English POS tagging with Cross-stitch | accuracy: 51.7±1.9 -> 52.1±1.0 | Helped |
| Replace the layered sub-encoder architecture with one recursive set of sub-encoders (ablation) | MT-RMN · Hindi-English POS tagging | F1: 53.4±2.4 -> 52.7±0.8 (NS-MTL); 21.4±3.2 -> 19.5±0.6 (Cross-stitch); 57.2±1.8 -> 56.6±0.7 (Sluice); 62.6±2.4 -> 61.3±1.5 | Hurt |
Closest papers by meaning
- Latent Multi-Task Architecture Learning (AAAI 2019 · similarity 0.88)
- End-To-End Multi-Task Learning With Attention (CVPR 2019 · similarity 0.87)
- Modular Universal Reparameterization: Deep Multi-task Learning Across Diverse Domains (NeurIPS 2019 · similarity 0.87)
- Breaking Neural Network Scaling Laws with Modularity (ICLR 2025 · similarity 0.85)
- Joint Scheduling of Causal Prompts and Tasks for Multi-Task Learning (CVPR 2025 · similarity 0.85)
- MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 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.