| <html><head><meta charset="utf-8"><style> | |
| body{margin:0;background:#f5f1e8;font:18px system-ui;color:#172554}.p{max-width:1100px;margin:auto;padding:36px}.h{background:#172554;color:white;padding:28px;border-radius:22px}.grid{display:grid;grid-template-columns:repeat(2,1fr);gap:18px;margin-top:18px}.c{background:white;padding:22px;border-radius:18px;border:1px solid #cbd5e1}.n{font-size:42px;font-weight:800;color:#c2410c}h1{margin:0 0 8px}h2{color:#0f766e}a{color:inherit;text-decoration:none}.pill{display:inline-block;background:#fed7aa;padding:6px 10px;border-radius:99px}</style></head><body><main class="p"><section class="h"><div class="pill">ICML 2026 reproduction</div><h1>Gaussian MFVI can overestimate predictive variance</h1><p>Exact conjugate-BLR calculations independently confirm the mechanism, calibration tradeoff, and high-dimensional rank-one pathology.</p></section><div class="grid"><a class="c" href="#/claim-1-mfvi-predictive-variance-overestimation-theorem-3-7"><h2>Expected ID variance</h2><div class="n">50 / 50</div><p>correlated-design trials had MFVI minus exact expected predictive variance ≥ 0.</p></a><a class="c" href="#/claim-2-calibration-forces-directional-variance-tradeoffs-lemma-3-6"><h2>Directional calibration</h2><div class="n">≤ 10⁻¹²</div><p>numerical error in the eigenbasis mean-ratio identity, with under- and over-estimated directions.</p></a><a class="c" href="#/claim-3-rank-one-pathology-approaches-prior-predictive-variance-section-5-1"><h2>Rank-one pathology</h2><div class="n">P = 1024</div><p>MFVI predictive variance approaches the prior along the informed direction.</p></a><a class="c" href="#/claim-4-temperature-scaling-corrects-id-and-ood-predictive-variance-differently"><h2>Temperature</h2><div class="n">cold ID · warm OOD</div><p>analytic variance matching reproduces the direction-dependent temperature effect.</p></a></div><p>Scope: deterministic NumPy reproduction plus source audit; the full RBF/UCI sweep was not rerun.</p></main></body></html> | |
Xet Storage Details
- Size:
- 2.04 kB
- Xet hash:
- 8ed6395db4eb7d868b3ae8d8df81f11b738858671e73f28f569003c01d2ddd62
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.