Research map / Architectures and efficiency / RNNs and normalization
Deciding How to Decide: Dynamic Routing in Artificial Neural Networks
Mason McGill, Pietro Perona
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 dynamic routing instead of static routing at a fixed computational budget | Image-classification networks · Hybrid MNIST/CIFAR-10 | accuracy | Improved |
| Use actor learning instead of critic learning | Dynamically-routed networks · Hybrid MNIST/CIFAR-10 | accuracy | Improved |
| Use optimistic critic learning instead of pragmatic critic learning (ablation) | Dynamically-routed networks · Hybrid MNIST/CIFAR-10 | accuracy | Hurt |
| Replace cross-entropy error with classification error in the pragmatic critic utility-regression term (ablation) | Pragmatic critic networks · Hybrid MNIST/CIFAR-10 | accuracy | Hurt |
| Add regularization on routing decisions (ablation) | Dynamically-routed networks · Hybrid MNIST/CIFAR-10 | peak accuracy | Hurt |
| Add a nonzero decisiveness cost to encourage exploration (ablation) | Actor networks · Hybrid MNIST/CIFAR-10 | training effectiveness | Improved |
| Apply throughput-adjusted learning rates (ablation) | Actor and critic networks · Hybrid MNIST/CIFAR-10, computational-resource-abundant high-accuracy contexts | performance | Improved |
| Use junctions with both 2 and 3 branches rather than only 2-way junctions (ablation) | Dynamically-routed networks · Hybrid MNIST/CIFAR-10 under tight computational constraints | accuracy | Improved |
| Use higher-degree branching (ablation) | Dynamically-routed networks · Hybrid MNIST/CIFAR-10 without computational constraints | overfitting | Hurt |
| Train one actor network conditioned on computation cost instead of separate networks for fixed costs (ablation) | Actor network · Hybrid MNIST/CIFAR-10 across varying computational constraints | memory consumption and training time: 8-fold reduction | Improved |
| Use dynamic routing instead of static routing | Actor networks · CIFAR-2 | peak accuracy | Improved |
| Use dynamic routing instead of static routing | Actor networks · CIFAR-10 | accuracy | Mixed |
Closest papers by meaning
- Adaptive Activation Thresholding: Dynamic Routing Type Behavior for Interpretability in Convolutional Neural Networks (ICCV 2019 · similarity 0.87)
- Runtime Network Routing for Efficient Image Classification (TPAMI 2018 · similarity 0.87)
- Differentiable Dynamic Wirings for Neural Networks (ICCV 2021 · similarity 0.86)
- Dynamic Neural Networks: A Survey (TPAMI 2021 · similarity 0.86)
- Exploring Randomly Wired Neural Networks for Image Recognition (ICCV 2019 · similarity 0.86)
- Three Mechanisms of Feature Learning in a Linear Network (ICLR 2025 · similarity 0.83)
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.
