Research map / Learning theory and algorithms / Prediction and decision losses
Approximation-Free Differentiable Oblique Decision Trees
Subrat Prasad Panda, Blaise Genest, Arvind Easwaran
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 |
|---|---|---|---|
| replace STE with annealed Top-k selection (ablation) | DTSemNet regression · CTSlice | RMSE: 1.45 -> 1.09 | Improved |
| replace STE with annealed Top-k selection (ablation) | DTSemNet regression · Pol | RMSE: 7.62 -> 6.61 | Improved |
| replace STE with annealed Top-k selection (ablation) | DTSemNet regression · Comp-Active | RMSE: 2.65 -> 2.48 | Improved |
| replace STE with annealed Top-k selection (ablation) | DTSemNet regression · PDBBind | RMSE: 1.34 -> 1.33 | No clear effect |
| replace STE with annealed Top-k selection (ablation) | DTSemNet regression · Regression Table 1 average | avg % distance to best: 8.8% -> 0.6% | Improved |
| update top-k regressors instead of only one (Top-4 vs Top-4/Reg-1) (ablation) | DTSemNet Top-k · CTSlice leaf utilization after epoch 1 | regressors with >100 samples: 16 -> 22 | Improved |
| update internal nodes with only top-k leaves instead of all leaves (Top-4/Reg-1 vs STE, end of training) (ablation) | DTSemNet Top-k · CTSlice leaf utilization at end | regressors used: 8 -> 15 | Improved |
| update top-k regressors and use top-k routing (Top-4 vs STE, end of training) (ablation) | DTSemNet · CTSlice leaf utilization at end | regressors with >180 samples: 8 -> 20 | Improved |
| use Top-2 instead of STE to learn the root weight with other nodes frozen (ablation) | DTSemNet · synthetic teacher-student root weight | convergence to teacher weight | Improved |
| replace DGT's STE architecture with DTSemNet | DGT · Dry Bean | accuracy: 89.0 -> 91.4 | Improved |
| replace DGT with DTSemNet | DGT · MNIST (height 8) | accuracy: 94.00 -> 96.16 | Improved |
| replace DGT with DTSemNet | DGT · Letter (height 10) | accuracy: 86.13 -> 89.19 | Improved |
Closest papers by meaning
- GradTree: Learning Axis-Aligned Decision Trees with Gradient Descent (AAAI 2024 · similarity 0.87)
- Learning Small Decision Trees with Few Outliers: A Parameterized Perspective (AAAI 2024 · similarity 0.87)
- Symbolic Regression Enhanced Decision Trees for Classification Tasks (AAAI 2024 · similarity 0.86)
- Branches: Efficiently Seeking Optimal Sparse Decision Trees via AO* (ICML 2025 · similarity 0.85)
- Regression Trees Know Calculus (NeurIPS 2025 · similarity 0.85)
- Leveraging Predictive Equivalence in Decision Trees (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.
