Research map / Architectures and efficiency / RNNs and normalization
Blockout: Dynamic Model Selection for Hierarchical Deep Networks
Check 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 |
|---|---|---|---|
| replace the Inception baseline with Blockout (hard, learned) | Inception with two fully connected layers · CIFAR-100 test | accuracy: 64.32 -> 65.66 | Helped |
| replace the Inception baseline with Blockout (hard, learned) | Inception with two fully connected layers · ImageNet validation | top-1 accuracy: 73.88 -> 74.83 | Helped |
| replace the Inception baseline with Blockout (soft, learned) | Inception with two fully connected layers · CIFAR-10 test | accuracy: 93.52 -> 94.41 | Helped |
| add random hard cluster selection to learned soft clustering (ablation) | Blockout · CIFAR-100 test | accuracy: 63.57 -> 65.66 | Helped |
| add random hard cluster selection to learned soft clustering (ablation) | Blockout · ImageNet validation | top-1 accuracy: 72.43 -> 74.83 | Helped |
| add random hard cluster selection to learned soft clustering (ablation) | Blockout · CIFAR-10 test | accuracy: 94.41 -> 93.55 | Hurt |
| learn cluster probabilities instead of fixing them at 0.5 (ablation) | Blockout with hard cluster assignments · CIFAR-100 test | accuracy: 65.62 -> 65.66 | No clear effect |
| learn cluster probabilities instead of fixing them at 0.5 (ablation) | Blockout with hard cluster assignments · ImageNet validation | top-1 accuracy: 74.44 -> 74.83 | Helped |
| learn cluster probabilities instead of fixing them at 0.5 (ablation) | Blockout with hard cluster assignments · CIFAR-10 test | accuracy: 93.67 -> 93.55 | Hurt |
| vary the number of clusters from 2 to 4 to 6 (ablation) | Blockout, CIFAR-100 with 512 nodes · CIFAR-100 test | accuracy: 64.54; 65.93; 65.66 | Mixed |
| vary the number of clusters from 2 to 6 to 15 (ablation) | Blockout, ImageNet with 4096 nodes · ImageNet validation | top-1 accuracy: 73.78; 74.83; 74.19 | Mixed |
| increase the number of nodes from 512 to 1024 to 2048 (ablation) | Blockout, CIFAR-100 with 6 clusters · CIFAR-100 test | accuracy: 65.66; 66.69; 66.71 | Helped |
Closest papers by meaning
- BlockDrop: Dynamic Inference Paths in Residual Networks (CVPR 2018 · similarity 0.88)
- DropBlock: A regularization method for convolutional networks (NeurIPS 2018 · similarity 0.87)
- Adaptive Dropout with Rademacher Complexity Regularization (ICLR 2018 · similarity 0.86)
- Hierarchical Selective Classification (NeurIPS 2024 · similarity 0.85)
- Hypernetworks for Dynamic Feature Selection (NeurIPS 2026 · similarity 0.84)
- How Deep Networks Learn Sparse and Hierarchical Data: the Sparse Random Hierarchy Model (ICML 2024 · similarity 0.84)
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.