backup: 2026-10-05 09:48

thesisintroduction.org

@@ -5,16 +5,16 @@
 #+title: thesis/introduction
 #+filetags: :project: :knowledge: :introduction:
 
-:PROPERTIES:
-:ID: thesis-introduction
-:END:
-
-#+title: thesis/introduction
-#+filetags: :project: :knowledge: :introduction:
-
 Chapter 1 of the BGU MSc thesis on ROLL. Written last, after Ch. 2–4 are stable.
 Parent: [[id:6294e2be-6189-4473-b363-a1dd9a75fb9b][thesis]]
 
+** Thesis Narrative (corrected 2026-08-29)
+
+The thesis is about learning under *label noise and class imbalance*, solved via
+direct AUC optimization. The hero method is ROLL-AoC. The surprise finding is
+*implicit* noise robustness — not a design goal, but a structural consequence of
+distributional score fitting.
+
 ** Guidelines
 
 - Target length: 4–6 pages
@@ -24,20 +24,33 @@ Parent: [[id:6294e2be-6189-4473-b363-a1dd9a75fb9b][thesis]]
 
 ** Structure
 
-1. *Hook* (~1–2 paragraphs) — concrete application where a wrong operating point is costly.
-   Candidates from existing prose: medical screening, fraud detection, astronomy/microscopy pipelines.
+1. *Hook* (~1–2 paragraphs) — concrete scenario where noisy labels + imbalance combine badly.
+   Good candidates from experiments: ChestX-ray14 (NLP-extracted labels, IR≈7.5:1),
+   credit card fraud (IR≈577:1), CIFAR-10N (crowdworker noise, 40% worst-annotator).
    Make the reader feel the problem before naming it.
-2. *Problem statement* (~1 paragraph) — class imbalance + why accuracy is the wrong metric +
-   the TPR@FPR objective in one crisp sentence.
-3. *Contributions* (bulleted list, ~5 items) — what ROLL does that prior work does not:
-   - Differentiable TPR@FPR loss (and FPR@TPR by label/score symmetry)
-   - Gaussian, Beta, KDE instantiations with closed-form gradients
-   - Gradient-balance property (updates independent of class-size ratio)
-   - Gradient-locality property (threshold-focused gradient mass)
-   - Empirical evaluation on KEEL benchmark + larger datasets
+
+2. *Problem statement* (~1–2 paragraphs) — three interlocking failures:
+   - Standard losses (BCE, focal) fail under imbalance (accuracy wrong metric) AND noise
+     (mislabeled samples produce large isolated gradient outliers)
+   - Noise-robust losses (GCE, MAE) stay per-sample — no AUC connection
+   - AUC-optimizing methods (LibAUC) use pairwise surrogates that lose direct connection to the curve
+   Gap: no single method simultaneously optimizes AUC, handles imbalance structurally, and is
+   robust to label noise.
+
+3. *Contributions* (bulleted list):
+   - ROLL: general framework for differentiable ROC-based optimization via score distribution fitting
+   - Three instantiations: Gaussian, Beta, KDE (KDE is primary)
+   - ROLL-AoC: direct numerical AUC integration via 99 per-FPR evaluations — distinct from pairwise surrogates
+   - Gradient balance property: updates invariant to class-size ratio
+   - Gradient locality property: gradient mass concentrates near operating threshold
+   - Empirical: *implicit* noise robustness — ROLL-AoC outperforms noise-designed losses (GCE, MAE)
+     at 20% noise, consistent across tabular (KEEL), image (CIFAR-10N), medical imaging (ChestX-ray14)
+
 4. *Roadmap* (~1 paragraph) — one sentence per chapter, matching final chapter titles exactly.
 
 ** Gotchas
 
 - Roadmap must exactly match final chapter titles — write it last
-- Don't use the inseparability / cascade-classifier tangent from the old draft; it dilutes the argument
+- Don't use the inseparability / cascade-classifier tangent from the old draft; it dilutes the argument
+- Don't frame as "operating point optimization" — the pivot is AUC + label noise
+- The noise robustness is *implicit* (emergent), not a design objective — this is the surprising result to sell