22042026 daily update .md

Updates/daily_updates/22042026-DailyUpdate.md

@@ -0,0 +1,26 @@
+**Daily Update — Bad Case Analysis, Improvement Report & Initial Experiments**
+*Image embedding benchmark · Tobacco-3482*
+
+**Completed**
+
+- Improved the bad case `.png` visualisations for better readability and repushed to HuggingFace. Added `bad_cases_grids.html` for improved grid-layout readability alongside the per-case PNGs.
+- Created `figures/bad_cases/bad_case_report.html` with per-label summary tables, confusion heatmaps, retrieval error rate bar charts, colour-coded error severity (green <25%, amber 25–50%, red >50%), root cause analysis, and a full improvement roadmap across four categories (preprocessing, architecture/pooling, training objectives, post-processing/retrieval).
+- Implemented and benchmarked **mean-centering** post-processing across SigLIP2 SO400M and ModernVBERT Bi — consistent gains across all metrics, most notably ModernVBERT Bi kNN@5 +4.73% (80.34% → 85.08%).
+- Implemented and benchmarked **CLS + mean-patch pooling** for SigLIP2 SO400M (dim=2304): kNN@1 +2.73% over CLS-only baseline (84.36% → 87.09%), with further gains when combined with centering (87.80% kNN@1, 89.81% kNN@5, 90.53% Clf Acc).
+- Ran **per-label kNN@1 error breakdown** across all six variants — CLS+patch+center is best overall and wins on 4 of 8 labels including Report (44.4% → 35.2%) and Note (35.6% → 25.4% tied with ModernVBERT Bi+ctr).
+- Implemented **query expansion** swept across k ∈ {1, 3, 5, 10} for all variants — marginal gains at best; CLS+patch+center with k=3 gives +0.57% kNN@1 (87.80% → 88.38%) but kNN@5 drops 2.15%. Query expansion does not reliably improve retrieval on this dataset.
+
+**In Progress**
+
+- Consolidating results — to be pushed to HuggingFace tomorrow.
+
+**Blockers**
+
+- Couldn't sync commit to huggingface
+
+**Key Learnings**
+
+- Centering consistently helps both models but ModernVBERT Bi benefits more than SigLIP2, suggesting its embedding space is less well-centred out of the box.
+- CLS+patch pooling trades classifier accuracy for kNN@1 accuracy on its own, but combining both with centering gives the best results across all three metrics — the two post-processing steps are complementary.
+- Query expansion does not help on Tobacco-3482 — gains on kNN@1 are marginal (≤0.57%) and kNN@5 consistently degrades, suggesting the neighbourhood structure is already noisy enough that averaging in neighbours introduces more error than it corrects.
+- CLS+patch+center is now the strongest single-model variant at 87.80% kNN@1 and 90.53% Clf Acc, but per-label breakdown shows Report (35.2%) and Note (25.4%) remain the hardest classes across all variants.