Research map / Architectures and efficiency / RNNs and normalization
Routing Networks: Adaptive Selection of Non-Linear Functions for Multi-Task Learning
Clemens Rosenbaum, Tim Klinger, Matthew Riemer
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 |
|---|---|---|---|
| use per-task hard routing trained with WPL instead of cross-stitch sharing | SimpleConvNet · CIFAR-MTL | accuracy: 60 -> 53 | Improved |
| use per-task hard routing trained with WPL instead of cross-stitch sharing | SimpleConvNet · MIN-MTL | accuracy: 57 -> 55 | Improved |
| use per-task hard routing trained with WPL instead of task-specific fully connected layers | SimpleConvNet · MNIST-MTL | accuracy: 99 -> 96 | Improved |
| use WPL rather than tabular Q-learning | CIFAR-MTL routing network · CIFAR-MTL | accuracy: 60 -> 55 | Improved |
| use multiple agents, one per task, rather than a single agent (ablation) | CIFAR-MTL routing network · CIFAR-MTL | accuracy: 60 -> 44 | Improved |
| use a tabular policy instead of an approximate policy (ablation) | CIFAR-MTL routing network · CIFAR-MTL | accuracy: 60 -> 25 | Improved |
| add a learned dispatcher to assign instances to agents (ablation) | CIFAR-MTL routing network · CIFAR-MTL | accuracy: 60 -> 41 | Hurt |
| replace hard routing with soft mixture routing (ablation) | CIFAR-MTL routing network · CIFAR-MTL | accuracy: 60 -> 35 | Hurt |
| allow non-layered routing with PASS (ablation) | CIFAR-MTL routing network · CIFAR-MTL | accuracy: 60 -> 46 | Hurt |
| add router-choice gradients to the function-block gradients (ablation) | routing network training · classification experiments | performance | No clear effect |
| use average historical block-choice probability rather than average historical choice count for the collaboration reward (ablation) | routing network router · routing experiments | final performance | No clear effect |
| set collaboration reward to zero or a small value (ablation) | routing network router · CIFAR-MTL and MIN-MTL | accuracy | Improved |
Closest papers by meaning
- Not All Tasks Are Equally Difficult: Multi-Task Deep Reinforcement Learning with Dynamic Depth Routing (AAAI 2024 · similarity 0.88)
- MTL-KD: Multi-Task Learning Via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver (NeurIPS 2025 · similarity 0.87)
- Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Memory-Enhanced Neural Solvers for Routing Problems (NeurIPS 2025 · similarity 0.86)
- SkillOrchestra: Learning to Route Agents via Skill Transfer (NeurIPS 2026 · similarity 0.85)
- Dynamic Latent Routing (NeurIPS 2026 · 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.
