Writing study: DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging
← NeurIPS 2026 oral writing corpus
Krishna Khadka, Yu Lei, Raghu Kacker, D. R Kuhn · Official NeurIPS 2026 oral record
Figure, table and section ledger
Editorial contribution type: Empirical analysis, diagnosis & auditing (primary).
Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.
- Canonical PDF
- 10 physical pages
- Before references / appendix
- ≈ 8.32 page-equivalents
- Main figures / tables
- 2 / 2
- Attached appendix figures / tables
- 0 / 0
- Attached appendix span
- No boundary detected; not assumed zero
Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.
Source-bound correction: Source-checked 2026-10-02: add the singular Acknowledgment heading on physical page 9. References start later on the same page. The 10-page author version has no appendix or checklist.
Section-by-section lengths and source pages
The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.
| Top-level heading | Starts on PDF page | Estimated span | Share of entire PDF |
|---|---|---|---|
| 1. Introduction | 1 | ≈ 1.19 pp. | 11.9% |
| 2. Related Work | 2 | ≈ 0.51 pp. | 5.1% |
| 3. Background | 3 | ≈ 0.45 pp. | 4.5% |
| 4. Motivation | 3 | ≈ 0.25 pp. | 2.5% |
| 5. Approach | 3 | ≈ 1.8 pp. | 18% |
| 6. Experimental Design | 5 | ≈ 2.91 pp. | 29.1% |
| 7. Results and Discussion | 8 | ≈ 0.26 pp. | 2.6% |
| 8. Limitations | 8 | ≈ 0.32 pp. | 3.2% |
| 9. Conclusion | 9 | ≈ 0.13 pp. | 1.3% |
| Acknowledgment | 9 | ≈ 0.11 pp. | 1.1% |
| References | 9 | ≈ 1.68 pp. | 16.8% |
All detected figure and table locations
Attached appendix
No recognized caption labels.
Argument and organization
- Problem and contribution
The paper frames explanation quality as preserving a model prediction with a sufficient subset of internal units, rather than requiring differentiable activation maps. Its key qualifier is 1-minimality: no remaining single removal preserves the stated condition, which is not a proof of globally smallest cardinality.
- Argument sequence
Background and motivation establish why gradient-based explanations can be problematic; the delta-debugging procedure then defines the intervention. The experimental section declares research questions on faithfulness and localization before listing datasets, architectures, baselines and metrics.
- Evidence and verification
The evaluation separates faithfulness from localization and reports computation time rather than treating all saliency maps as equivalent explanations. Table 1 groups architecture families; Table 2 includes localization precision/recall and region counts. The ImageNet and chest-X-ray experiments probe different settings and should not be collapsed into a clinical validity claim.
Problem, claim and contribution
The paper frames explanation quality as preserving a model prediction with a sufficient subset of internal units, rather than requiring differentiable activation maps. Its key qualifier is 1-minimality: no remaining single removal preserves the stated condition, which is not a proof of globally smallest cardinality.
Abstract and introduction, pp. 1–2; method definition and Algorithm 1, pp. 4–5 · Main paper
Argument and section order
Background and motivation establish why gradient-based explanations can be problematic; the delta-debugging procedure then defines the intervention. The experimental section declares research questions on faithfulness and localization before listing datasets, architectures, baselines and metrics.
Background/motivation, p. 3; method pp. 4–5; RQ1/RQ2 and evaluation setup, pp. 5–7 · Main paper
Figures, captions and visual explanation
Algorithm 1 gives explicit inputs, output and line-numbered complement tests, making the discrete search easier to inspect than a prose-only recipe. Figure 1 uses a repeated image-by-method saliency grid across architecture rows. The grid is informative but its dense small metric text over colored images is a readability warning at publication scale.
Algorithm 1, p. 5; Figure 1, p. 7; rendered pages 5 and 7 · Main paper
Tables, comparisons and evidence
The evaluation separates faithfulness from localization and reports computation time rather than treating all saliency maps as equivalent explanations. Table 1 groups architecture families; Table 2 includes localization precision/recall and region counts. The ImageNet and chest-X-ray experiments probe different settings and should not be collapsed into a clinical validity claim.
Tables 1–2 and evaluation protocol, pp. 5–7; qualitative X-ray results, Figure 2, p. 8 · Main paper
Appendix and reproducibility
Because the available version has no appendix, the query cost, interaction between units and access assumptions need to remain visible in its main discussion. A future supplement could document additional cases, but its absence must not be filled with invented appendix evidence.
Discussion/limitations and conclusion, pp. 8–9; references pp. 9–10 · Main paper
Clarity and writing conventions
The RQ1/RQ2 structure makes the experimental purpose unusually easy to locate. Precise terms such as sufficient and 1-minimal need their definitions nearby, because casual wording such as smallest explanation would claim more than the procedure establishes.
Method definition pp. 4–5; research questions p. 5; limitations pp. 8–9 · Main paper
What to reuse—and what not to infer
Transferable lesson: Introduce a discrete explanation algorithm with explicit pseudocode, then evaluate separately the properties named in its definition; use qualitative maps to illustrate those tests, not replace them.
Boundary: This is a public two-column author version, not verified NeurIPS camera-ready formatting. Prediction-preserving saliency and box overlap do not establish causal explanations or clinical diagnostic validity.
Apply this lesson using the argument and evidence worksheet.
Inspection record and sources
- Checked
- 2026-10-02
- Reading status
- Writing-focused manuscript scan + selected visuals inspected
- Main text
- Writing-oriented scan across manuscript framing, section sequence, figure/table captions, selected argument/evaluation prose and conclusion; not exhaustive technical or proof verification. Full 10-page author manuscript scanned; main argument and experiments pp. 1–9, references pp. 9–10. Rendered algorithm/evaluation setup p. 5 and qualitative comparison p. 7. Source is the exact public author version listed below, not a verified accepted/camera-ready revision.
- Appendix / supplement
- Writing-oriented appendix/documentation scan: No separate appendix is present in this public 10-page version. Method, experiments and limitations are in the main document. Source code, data, model outputs and external video assets were not executed or independently validated.
- Visually inspected PDF pages
- 5, 7
This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.