The negative results
Matched comparison: same count, same Adam steps, same images; the baseline is direct per-image k-means plus the same refinement.
| Result | Numbers | Why it failed |
|---|---|---|
| Split growth (s6, 16×16, 200 img) | At 64: 0.0000866 vs k-means 0.0000622 (+39%, 2 steps/round); 0.0000776 vs 0.0000585 (+33%, 4 steps) | Good layouts are not nested; overlapping children add little to the basis; greedy choices are myopic |
| Split criteria (s6) | Residual 0.0000866, mass 0.0000925, info gain 0.0000937, random 0.0000955; oracle (50 img) no better than residual | Entropy counts distinct codes, not error (count follows code entropy 0.89, error 0.25); oracle is one-step myopic |
| Merge-down from 256 (s7, 100 img) | +29 / 21 / 17 / 11% at 128 / 64 / 32 / 16 (input512); +17 / 17 / 14 / 12% (shift8) | Merge score ranks pairs poorly (rank correlation 0.57 best, 0.29 shift8); non-nesting in reverse |
| Adaptive count, mean 64 (s7, 100 img) | −1.2% oracle, −1.0% proxy (input512); −0.9%, −0.5% (shift8); 74 of 100 images kept 64 | Per-image error curves nearly parallel (slope 0.85 ± 0.11 per doubling); ideal bound about 2.4% |
Caveats: plain PyTorch quick experiments; 50 to 200 images; one run; settings tuned on the scored images; no learned split or merge network trained.
Consequence: one model serving any count by direct clustering, not a split or merge tree.
Defence: “Hierarchical placement was tested at matched budget before building it. Splitting, merging and per-image counts lost by 11 to 39% or gained about 1%, so the model clusters directly at whatever count is asked for.”