Conversation
related_work.md: Added ABR accuracy numbers alongside each paper, FEAT direct comparison (+1.85% vs FEAT Conv4), ProtoDiff comparison, ResNet-12 failure mode analysis, cross-domain degradation, and 2024-2025 papers to monitor (NovPro, VPT, UniSiam). related_work_table.md: Added four accuracy tables — Conv4 mini, ResNet-12 mini, CIFAR-FS, tieredImageNet, and cross-domain — with real numbers from experiments. Summary positioning table added at end. sota_landscape.md: Updated Tier 3 (Conv4) and added Tier 2b (ResNet-12) rows with correct 60k episode numbers instead of old 10k numbers. publication_roadmap.md: Marked all experiments as complete, converted required experiments list to writing tasks, updated selling points. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
paper/paper.md contains the complete paper: - Abstract, Introduction, Related Work, Method, Experiments, Analysis, Conclusion, References - All numerical results drawn directly from experiment CSVs (no hallucination) - 5 figures referenced from results/publication/ - 18 real references (NeurIPS, CVPR, ICLR, ICML, ECCV) - Honest documentation of both positive results and failure modes - Paired bootstrap significance testing reported throughout Key headline results in paper: - ABR-MLP 57.00% miniImageNet 1-shot (vs FEAT 55.15%, p<0.001) - CIFAR-FS 1-shot +2.64% (p<0.001) - ResNet-12 1-shot failure mode documented - Cross-domain (mini→CUB) degradation documented Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Audit found the original ABR results were trained jointly with the backbone (not the frozen plug-in the paper claims) and the ResNet-12 1-shot baseline was undertrained (15k vs 40k). Re-ran every condition as a true plug-in (freeze trained baseline backbone, train only ABR) and corrected the ResNet-12 baseline. Findings (plug-in, frozen backbone): - ABR-MLP helps only at 1-shot on weak backbone: mini +1.44%, CIFAR-FS +3.10% - ResNet-12 (corrected 40k baseline): neutral (within CI) - Cross-domain mini->CUB: neutral; the joint -3.64% collapse was an artifact - GNN/Attn and all 5-shot: neutral Changes: - train_plugin.py: frozen-backbone plug-in trainer - scripts/run_conv4_plugin.sh, run_resnet_plugin.sh: plug-in batch runners - scripts/run_crossdomain.py: support plug-in variants + run-name override - src/training/trainer.py: optimiser only sees trainable params (frozen-safe) - results/plugin/: consolidated plug-in CSVs (summary + per-condition + cross-domain) - results/resnet12_40k_check/: corrected ResNet-12 baseline - RESULTS.md: joint vs plug-in side-by-side tables - paper/paper_accv.md: thesis updated to honest plug-in findings - DEBUG_RESNET12.md: audit notes Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…shot results Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Remove 475 intermediate result files (per-condition CSVs, UMAP/boundary figures, confusion matrices) — only publication-ready figures kept in results/publication/ (11 figures, PDF+PNG) - Remove internal planning docs (ACCV_ACTION_PLAN, AGENTS, DEBUG_RESNET12, experiment_log, research_notes, docs/, paper/, publication_plan/) - Remove pptx presentations and drawio source files - Update .gitignore to cover all of the above permanently - Rewrite README.md — paper-facing: key results table, publication figures directory, quick-start, architecture summary, citation block - Rewrite RESULTS.md — plug-in protocol only, clean tables with 95% CI, full significance table (24 conditions), multi-seed, cross-domain, geometry; removes superseded 10k runs and internal TODO sections - Add abr_variants.html — interactive diagram of all three ABR variants - Add abr_pseudocode.txt — pseudocode for all three refinement algorithms Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Remove old tracked names (fig1_main_results, fig3_geometry, etc.) and add current paper_fig1–11 set. Add generate_figures.py to publication dir. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Embed pipeline, variants, accuracy gain, when-ABR-works, geometry, multi-seed, and all-variants figures inline. Table of figures becomes a linked index. Architecture section moved above results. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Keep: when_abr_works, accuracy_gain, radar_geometry. Remove: pipeline, variants, geometry correlation, multiseed, all_variants. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Subtitle: "Under review, ACCV 2026" (not accepted) - Remove citation block (inappropriate before acceptance) - Rename "Publication Figures" -> "Figures" - Remove "ACCV paper figures" label from script comments Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.