Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Tractable Approximate Gaussian Inference for Bayesian Neural Networks
James-A. Goulet, Luong Ha Nguyen, Saeid Amiri
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 |
|---|---|---|---|
| use full instead of diagonal hidden-unit covariance (ablation) | TAGI, 1D toy · 1D toy regression y=x^3+v | propagation of input uncertainty: diagonal covariances unable to retrieve reference std 6x^2/10; full covariance follows it closely | Improved |
| increase hidden units per layer from 5 to 50 (ablation) | TAGI with full covariance · 1D toy regression | posterior correlation between parameters: non-zero correlations from extensive to a small fraction of the covariance matrix | Improved |
| use diagonal instead of full covariance with many hidden units (>100) (ablation) | TAGI · sum-of-units toy model | posterior correlation of hidden units: non-negligible for <10 units, almost zero for >100 | No clear effect |
| increase MNIST hidden units from 100 to 800 (B=1) (ablation) | TAGI · MNIST test | classification error (%): 2.29 -> 1.54 | Improved |
| increase MNIST hidden units from 100 to 800 (B=10) (ablation) | TAGI · MNIST test | classification error (%): 2.45 -> 1.53 | Improved |
| train to early-stopped epoch instead of one epoch (A=800, B=1) (ablation) | TAGI · MNIST test | classification error (%): 3.15 -> 1.54 | Improved |
| train to early-stopped epoch instead of one epoch (A=100, B=10) (ablation) | TAGI · MNIST test | classification error (%): 3.56 -> 2.45 | Improved |
| use batch size 10 instead of 1 (A=800) (ablation) | TAGI · MNIST test | classification error (%): 1.54 -> 1.53 | No clear effect |
| use batch size 10 instead of 1 (A=100) (ablation) | TAGI · MNIST test | classification error (%): 2.29 -> 2.45 | Hurt |
| compare TAGI with VMG on Energy | TAGI · Energy (UCI regression) | RMSE: 1.46 vs 0.54 (VMG) | Hurt |
| compare TAGI with VMG on Concrete | TAGI · Concrete (UCI regression) | RMSE: 5.72 vs 4.89 (VMG) | Hurt |
| compare TAGI with MC-Dropout on Boston | TAGI · Boston (UCI regression) | RMSE: 2.98 vs 2.97 (MC-Dropout) | No clear effect |
Closest papers by meaning
- Efficient Low Rank Gaussian Variational Inference for Neural Networks (NeurIPS 2020 · similarity 0.88)
- py/cuTAGI: An Open-Source Library for Tractable Approximate Gaussian Inference in Bayesian Neural Networks (JMLR 2026 · similarity 0.87)
- A Study of Bayesian Neural Network Surrogates for Bayesian Optimization (ICLR 2024 · similarity 0.86)
- Do Bayesian Neural Networks Actually Behave Like Bayesian Models? (ICML 2025 · similarity 0.86)
- Rethinking Approximate Gaussian Inference in Classification (NeurIPS 2025 · similarity 0.85)
- Connecting the Dots: Is Mode-Connectedness the Key to Feasible Sample-Based Inference in Bayesian Neural Networks? (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.
