logbook.json

{
  "schema_version": 1,
  "title": "On the Power of Source Screening for Learning Shared Feature Extractors",
  "emoji": "🔍",
  "space_id": "snaykey/repro-source-screening-shared-feature-extractors",
  "paper": {
    "arxiv_id": "2602.16125",
    "openreview_id": "dTMrITkTr5"
  },
  "tags": [
    "icml2026-repro",
    "paper-dTMrITkTr5"
  ],
  "updated_at": "2026-07-24T19:42:38+00:00",
  "root": {
    "slug": "index",
    "title": "On the Power of Source Screening for Learning Shared Feature Extractors",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-minimax-optimal-rate",
        "title": "Theorem 2 shows that training on an admissible subset of sources achieves the minimax-optimal statistical rate O(√(d/(Nλ_k))) for shared low-dimensional subspace estimation (Theorem 2).",
        "file": "pages/claim-1-minimax-optimal-rate/page.md",
        "children": []
      },
      {
        "slug": "claim-2-spectral-norm-condition",
        "title": "Theorem 3 establishes that a 'good' admissible source subpopulation exists whenever a spectral-norm condition on the source covariates holds (Theorem 3).",
        "file": "pages/claim-2-spectral-norm-condition/page.md",
        "children": []
      },
      {
        "slug": "claim-3-algorithm-correctness",
        "title": "Algorithm 1 (genie-aided) and Algorithm 2 (empirical) for identifying informative source subsets are proven to output admissible subsets with high probability (Theorem 5, Algorithm 1, Algorithm 2).",
        "file": "pages/claim-3-algorithm-correctness/page.md",
        "children": []
      },
      {
        "slug": "claim-4-real-world-classification",
        "title": "On real-world evaluations, training with a screened source subset yields 74.2% classification accuracy on ACSIncome and 90.5% on CelebA smile classification (Table 2).",
        "file": "pages/claim-4-real-world-classification/page.md",
        "children": []
      },
      {
        "slug": "claim-5-figure-1-screened-vs-full",
        "title": "Figure 1 shows subspace reconstruction error is lower when training on a balanced, screened subset of sources than on the full heterogeneous source population (Figure 1).",
        "file": "pages/claim-5-figure-1-screened-vs-full/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      },
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 4595,
  "revision": "1784922158165495400"
}