Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Differentiable Compositional Kernel Learning for Gaussian Processes
Shengyang Sun, Guodong Zhang, Chaoqi Wang, Wenyuan Zeng, Jiaman Li, Roger Grosse
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 NKN instead of GP-RBF | Gaussian-process regression · UCI random-split benchmarks: Boston, Energy, Pow. Plant, Wine, Yacht | Test RMSE: Boston: 2.753±0.137 -> 2.506±0.150; Energy: 0.471±0.013 -> 0.254±0.020; Pow. Plant: 3.014±0.068 -> 2.675±0.074; Wine: 0. | Improved |
| Use NKN instead of GP-RBF | Gaussian-process regression · UCI random-split benchmarks: Boston, Concrete, Energy, Kin8nm, Naval, Pow. Plant, Wine, Yacht | Test log-likelihood: Boston: -2.434±0.069 -> -2.394±0.080; Concrete: -2.948±0.025 -> -2.842±0.263; Energy: -0.673±0.035 -> -0.213±0.162; Kin8 | Mixed |
| Use NKN instead of GP-SM4 | Gaussian-process regression · UCI random-split benchmarks: Boston, Energy, Pow. Plant, Wine, Yacht | Test RMSE: Boston: 2.979±0.162 -> 2.506±0.150; Energy: 0.316±0.018 -> 0.254±0.020; Pow. Plant: 2.781±0.071 -> 2.675±0.074; Wine: 0. | Improved |
| Use NKN instead of GP-SM4 | Gaussian-process regression · UCI random-split benchmarks: Concrete, Energy, Kin8nm, Naval, Pow. Plant, Wine, Yacht | Test log-likelihood: Concrete: -2.662±0.053 -> -2.842±0.263; Energy: -0.320±0.089 -> -0.213±0.162; Kin8nm: 1.387±0.006 -> 1.291±0.006; Naval: | Mixed |
| Use NKN instead of GP-SM | Gaussian-process regression · PCA-split regression benchmarks | Test RMSE: Boston: 8.600 -> 4.658; Concrete: 7.591 -> 6.242; Energy: 0.447 -> 0.459; Kin8nm: 0.065 -> 0.086; Naval: 0.000 -> 0.000; | Mixed |
| Use NKN instead of GP-SM | Gaussian-process regression · PCA-split regression benchmarks | Test log-likelihood: Boston: -4.404 -> -3.557; Concrete: -3.285 -> -3.112; Energy: -0.564 -> -0.649; Kin8nm: 1.322 -> 1.057; Naval: 9.037 -> | Mixed |
| Replace heuristic sum kernel with compositional NKN | Airline time-series GP · Airline extrapolation | extrapolation quality | Improved |
| Replace Automatic Statistician with NKN | Airline time-series kernel learning · Airline training time | training time (seconds): 6147 seconds -> 201 seconds | Improved |
| Use NKN instead of SM kernel | Texture extrapolation · Metal thread plate and paved pattern | visual reconstruction quality | Improved |
| Use NKN instead of generic RBF kernel | Bayesian optimization with Expected Improvement · Three additive toy-function benchmarks | best function value | Improved |
Closest papers by meaning
- Learning Compositional Sparse Gaussian Processes with a Shrinkage Prior (AAAI 2021 · similarity 0.88)
- Neural Kernels Without Tangents (ICML 2020 · similarity 0.88)
- Graph Neural Network-Inspired Kernels for Gaussian Processes in Semi-Supervised Learning (ICLR 2023 · similarity 0.87)
- Sparse Spectrum Warped Input Measures for Nonstationary Kernel Learning (NeurIPS 2020 · similarity 0.86)
- Scalable Levy Process Priors for Spectral Kernel Learning (NeurIPS 2017 · similarity 0.86)
- gp2Scale: A Class of Compactly Supported Non-Stationary Kernels and Distributed Computing for Exact Gaussian Processes on 10 Million Data Points (ICML 2026 · similarity 0.86)
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.
