| { |
| "schema_version": 1, |
| "title": "Reproduction: Expectation Consistency Loss (calibration under covariate shift)", |
| "emoji": "🎯", |
| "space_id": "snaykey/repro-expectation-consistency", |
| "paper": { |
| "openreview_id": "gFPPTokv9C" |
| }, |
| "tags": [ |
| "icml2026-repro", |
| "paper-gFPPTokv9C" |
| ], |
| "updated_at": "2026-08-02T00:00:00+00:00", |
| "root": { |
| "slug": "index", |
| "title": "Reproduction: Expectation Consistency Loss (calibration under covariate shift)", |
| "file": "pages/index.md", |
| "children": [ |
| { |
| "slug": "claim-1-thm31-ec-condition", |
| "title": "Theorem 3.1 establishes the Expectation Consistency Condition: canonical calibration under covariate shift holds for class k if and only if the conditional expectation of P(Y_k=1|X) matches between source and target distributions given the calibrated score S, which is strictly weaker than requiring full distributional alignment (Section 3.1, Theorem 3.1).", |
| "file": "pages/claim-1-thm31-ec-condition/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-2-thm32-sample-complexity", |
| "title": "Theorem 3.2 proves the Expectation Consistency Loss (ECL) estimator has O(B/epsilon^2) sample complexity with an explicit finite-sample bound on the gap between empirical and population ECL, matching the complexity of histogram-binning-based ECE estimation (Section 3.4, Theorem 3.2).", |
| "file": "pages/claim-2-thm32-sample-complexity/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-3-thm33-minibatch-unbiased", |
| "title": "Theorem 3.3 shows the mini-batch ECL formulation using auxiliary variables u_j^s and u_j^t yields an unbiased gradient estimator, E[∇_θ L̂_ecl^mini] = ∇_θ L̂_ecl, enabling standard SGD training (Section 3.5, Theorem 3.3).", |
| "file": "pages/claim-3-thm33-minibatch-unbiased/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-4-variants-softbin", |
| "title": "ECL is extended to top-label calibration (Appendix D.1, Section E.1) and class-wise calibration (Appendix D.2, Section E.2), with differentiable soft-binning implementations for canonical, class-wise, and top-label variants given in Appendix F.", |
| "file": "pages/claim-4-variants-softbin/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-5-table2-toplabel", |
| "title": "On MNIST-to-SVHN top-label calibration with LeNet-5, ECL reduces ECE to 21.5% versus 61.9% for the uncalibrated baseline; with ResNet20, ECL achieves 36.8% ECE versus 48.2% for PseudoCal; with DenseNet40, ECL achieves 38.4% ECE versus 80.8% uncalibrated (Table 2).", |
| "file": "pages/claim-5-table2-toplabel/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "claim-6-table1-reliability", |
| "title": "Table 1 shows ECL is the only compared method that simultaneously handles covariate shift, supports class-wise and canonical calibration, tolerates unbounded density ratios, and is mini-batch trainable, while Figure 2 reliability diagrams and Appendix Table 3 (source-vs-target calibration gap) further corroborate the approach on simulated data and the PACS dataset (Table 1, Figure 2, Appendix Table 3).", |
| "file": "pages/claim-6-table1-reliability/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "executive-summary", |
| "title": "executive-summary", |
| "file": "pages/executive-summary/page.md", |
| "children": [] |
| }, |
| { |
| "slug": "conclusion", |
| "title": "conclusion", |
| "file": "pages/conclusion/page.md", |
| "children": [] |
| } |
| ] |
| } |
| } |
