Add per-parameter pruning rows + joint pruning result. R1 remains leading Pareto point.

eval

dim20 joint pruned from R1: 17,129 nonzero, 3.6 mAP. Whole-head magnitude pruning.

eval

dim20 R1 + cls_weight pruned: 17,406 nonzero, 3.5 mAP. Greedy-stack on top of R1.

eval

dim20 reg_out pruned: 8 of 64 weights zeroed (12%), 3.8 mAP

eval

dim20 reg_hidden pruned: 55 of 320 weights zeroed (17%), 3.8 mAP

eval

dim20 ctr_weight pruned: 18 of 20 weights zeroed (90%), 3.7 mAP

eval

dim20 cls_weight pruned: 596 of 1600 weights zeroed (37%), 3.8 mAP

Add dim20 pruned R1/R2 rows: 18K nonzero params at 3.9 mAP, the most parameter-efficient detector in the table

dim20 project.weight pruning Pareto front from rigorous mAP-driven pruner

dim20 pruned R2 pycocotools eval: 3.8 mAP, 14.5 mAP@0.50

dim20 pruned R2: 17,988 nonzero params (26.6% of project.weight zeroed), 3.8 mAP

dim20 pruned R1 pycocotools eval: 3.9 mAP, 14.6 mAP@0.50

dim20 pruned R1: 18,121 nonzero params (25.7% of project.weight zeroed), 3.9 mAP retained

Delete legacy outputs/ tree (duplicates of heads/cofiber_threshold/ canonical files)

Add dim15 row: 17,751 params, 3.0 mAP — fills the dim20 to dim10 portion of the curve

dim15 pycocotools eval: 3.0 mAP, 11.5 mAP@0.50

dim15: 17,751 params, 3.0 mAP — fills in the dim20→dim10 portion of the curve

Add dim10 row: 13,426 params, 1.5 mAP — smallest 80-class detector above the 1.0 mAP threshold

dim10 pycocotools eval: 1.5 mAP, 5.6 mAP@0.50

dim10: 13,426 params, 1.5 mAP — floor candidate above the 1.0 threshold

Add dim5 row: 9,101 params, 0.3 mAP — 5 dimensions sit below the 80-class capacity floor

dim5 pycocotools eval: 0.3 mAP, 1.3 mAP@0.50

dim5: 9,101 params, 0.3 mAP. The 5-dim bottleneck is below the intrinsic capacity for 80-class separation.

Add training recipe table and checkpoint locations

Add base_model: facebook/EUPE-ViT-B and link backbone in prose

Update COCO mAP table with dim20 result (3.9 mAP, 22K params)

dim30: 768->30->80 bottleneck, 30,726 params, 80% SVD energy retention. Not yet trained.

Remove duplicate dim20_22k folder, keep dim20_20k with trained weights

dim20: 3.9 mAP from 22K params — 80-class COCO detector via SVD dimension selection

dim20: 3.9 mAP from 22K params. 768->20->80 bottleneck, SVD-initialized from pruned prototypes.

Update README with pruning curve and dim20, fix person detector ext4 path

Pruned R3: 46K nonzero, mAP 5.1 (dropped from 5.9). R2 at 62K is the sweet spot.

Add pruned_62k checkpoint (5.9 mAP, 30K weights zeroed, mAP improved), fix pruner naming

Clarify screening was on 10 val images, list all 25 heads, separate screened from unscreened

Fix stale eval protocol (10 vs 1000 images), add pruned result to COCO table

box32 pruned to 76K nonzero params: 5.7 mAP preserved, 15K weights zeroed

Add COCO mAP results: linear_70k 4.0 mAP, box32_92k 5.7 mAP (pycocotools, 5000 val)

box32_92k: 5.7 mAP from 92K params (43% improvement over linear_70k)

Reorganize cofiber_threshold: subfolders per run (linear_70k, box32_92k), remove root training scripts

Add untrained weights alongside trained

Rename artifacts to descriptive names, delete head_final

Cofiber Threshold: trained weights, COCO mAP 4.0 from 70K params, eval script

Add tropical, compression, curvature, renormalization heads

Add novel heads: Optimal Transport (det), Info Bottleneck (seg), Harmonic Depth (depth)