Research map / Architectures and efficiency / RNNs and normalization
Adaptive Neural Trees
Ryutaro Tanno, Kai Arulkumaran, Daniel Alexander, Antonio Criminisi, Aditya Nori
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 |
|---|---|---|---|
| Combine learned feature transformations with hierarchical router-based data partitioning (ablation) | Adaptive Neural Trees · SARCOS, MNIST, CIFAR-10 | prediction error: SARCOS: 1.38 vs 2.51 and 2.12; MNIST-A: 0.64 vs 0.74 and 3.18; CIFAR10-A: 8.31 vs 9.29 and 39.29 | Improved |
| Add router-based hierarchical partitioning (ablation) | Adaptive Neural Trees · SARCOS, MNIST-A/B/C, CIFAR10-A/B/C | prediction error: 1.38 -> 2.51; 0.64 -> 0.74; 0.72 -> 0.80; 1.62 -> 3.71; 8.31 -> 9.29; 9.15 -> 11.08; 9.31 -> 11.61 | Improved |
| Add learned transformer modules for feature learning (ablation) | Adaptive Neural Trees · SARCOS, MNIST-A/B/C, CIFAR10-A/B/C | prediction error: 1.38 -> 2.12; 0.64 -> 3.18; 0.72 -> 4.63; 1.62 -> 5.70; 8.31 -> 39.29; 9.15 -> 43.09; 9.31 -> 48.59 | Improved |
| Use single-path instead of multi-path inference | Adaptive Neural Trees · SARCOS, MNIST, CIFAR-10 | prediction error: 1.384 -> 1.542; 0.64 -> 0.69; 8.31 -> 8.32 | No clear effect |
| Use ANT-MNIST-B with smaller convolution kernels than ANT-MNIST-A | ANT-MNIST-A · MNIST | classification error: 0.64 -> 0.72 | Hurt |
| Match the root-to-path CNN's parameter count to ANT-MNIST-C | ANT-MNIST-C · MNIST | classification error: 1.68 -> 3.55 | Hurt |
| Initialize routers and transformers from a pre-trained single-path CNN | ANT-MNIST-A · CIFAR-10 | classification error: 8.31 -> 6.72 | Improved |
| Globally refine all modules after architecture growth | Adaptive Neural Trees · CIFAR-10 | test accuracy | Improved |
| Set growth-phase patience to 1 (ablation) | ANT-CIFAR10-C · CIFAR-10 | validation accuracy: 80% | Hurt |
| Set growth-phase patience to 15 (ablation) | ANT-CIFAR10-C · CIFAR-10 | validation accuracy: 87% | Hurt |
| Set growth-phase patience to 5 (ablation) | ANT-CIFAR10-C · CIFAR-10 | validation accuracy: 91% | Improved |
| Ensemble 8 ANT models | Adaptive Neural Trees · MNIST, CIFAR-10, SARCOS | prediction error: MNIST 0.64 -> 0.29; CIFAR-10 8.31 -> 7.76; SARCOS 1.384 -> 1.226 | Improved |
Closest papers by meaning
- End-to-End Learning of Decision Trees and Forests (IJCV 2019 · similarity 0.86)
- Learning concise representations for regression by evolving networks of trees (ICLR 2019 · similarity 0.86)
- Analyzing the tree-layer structure of Deep Forests (ICML 2021 · similarity 0.86)
- Training Decision Trees as Replacement for Convolution Layers (AAAI 2020 · similarity 0.85)
- Generative Model for Decision Trees (AAAI 2024 · similarity 0.84)
- Branches: Efficiently Seeking Optimal Sparse Decision Trees via AO* (ICML 2025 · 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.
