Research map / Learning theory and algorithms / Prediction and decision losses
Diminishing Returns Shape Constraints for Interpretability and Regularization
Maya Gupta, Dara Bahri, Andrew Cotter, Kevin Canini
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 |
|---|---|---|---|
| add convexity constraint (ablation) | calibrated linear model · Car Sales | Test MSE: 10727 -> 10593 | Improved |
| add diminishing-returns shape constraints (ablation) | calibrated linear model · Puzzle Sales | Test MSE: 8270 -> 8189 | Improved |
| add diminishing-returns shape constraints (ablation) | RTL · Puzzle Sales | Test MSE: 8838 -> 8315 | Improved |
| add concavity constraint (ablation) | RTL · Domain Pricing | Test MSE: 0.1109 -> 0.1078 | Improved |
| add concavity constraint (ablation) | calibrated linear model · Domain Pricing | Test MSE: 0.1150 -> 0.1120 | Improved |
| add concavity constraint (ablation) | RTL and calibrated linear · Wine Quality | Test MSE: 4.85 -> 4.83 | No clear effect |
| add concavity constraint on relatedness (ablation) | calibrated linear model · Query-Result Matching (1/10 train) | Test MSE: 0.710 -> 0.702 | Improved |
| add concavity constraint on relatedness (ablation) | calibrated linear model · Query-Result Matching (full train) | Test MSE: 0.701 -> 0.722 | Hurt |
| add shape constraints (monotonic and concave) over monotonic-only (ablation) | RTL · Query-Result Matching (full train) | Test MSE: 0.654 -> 0.654 | No clear effect |
| add monotonic constraints to unconstrained RTL calibrators (ablation) | RTL · Query-Result Matching (full train) | Test MSE: 0.639 -> 0.654 | Hurt |
| use shape-constrained lattice models instead of unconstrained DNN (ablation) | lattice models · Query-Result Matching (1/10 train) | Test MSE: test MSE std .002 (RTL) vs 1.69 (DNN) | Improved |
| use SCNN diminishing returns instead of concave-only (ablation) | SCNN · Query-Result Matching (1/10 train) | Test MSE: 0.673 -> 0.680 | Hurt |
Closest papers by meaning
- Multidimensional Shape Constraints (ICML 2020 · similarity 0.85)
- Loss Shaping Constraints for Long-Term Time Series Forecasting (ICML 2024 · similarity 0.83)
- How DNNs break the Curse of Dimensionality: Compositionality and Symmetry Learning (ICLR 2025 · similarity 0.82)
- An Infinite-Width Analysis on the Jacobian-Regularised Training of a Neural Network (ICML 2024 · similarity 0.82)
- The Star Geometry of Critic-Based Regularizer Learning (NeurIPS 2024 · similarity 0.82)
- Sharp feature-learning transitions and Bayes-optimal neural scaling laws in extensive-width networks (NeurIPS 2026 · similarity 0.82)
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.
