The amortisation gap
Two gaps (Cremer et al., 2018): the approximation gap is what the best placement of K Gaussians still misses (more Gaussians shrink it); the amortisation gap is the network's error minus the best per-image error (a better predictor shrinks it).
Amos's vocabulary (Amos, 2022): fully vs semi-amortised (does the model access the objective?); regression- vs objective-based learning. Model J is semi-amortised, objective-based: 16 Adam steps on the true error inside the network, trained on dense error. Precedents: semi-amortised VAEs, iterative amortised inference.
| 16×16, 64 Gaussians | Dense MSE | × best | Gap closed |
|---|---|---|---|
| Gridless64 one shot | 0.000104 | 2.24 | 0% |
| Clustering B | 0.0000804 | 1.73 | 41% |
| J, 16 steps | 0.0000587 | 1.26 | 79% |
| J, 32 steps | 0.0000531 | 1.14 | 88% |
| Best per-image (600 Adam, grid start, 100 imgs) | 0.0000465 | 1 | 100% |
The optimum is only the best found (non-convex; 600 steps give 0.0000465 from a grid, 0.0000558 from the Gridless64 layout), so the measured gap is a lower bound.
32×32 (input512), same first 100 images: k-means 0.000149; k-means + 16 steps 0.000117; J 16 steps 0.000112 (1.13×); J 32 steps 0.000104 (1.05×); k-means + 200 steps 0.0000988. The roadmap's “15%” divided an 800-image value by a 100-image one; matched it is 13% and 5%.
Defence sentence: “On the 32×32 DINOv3 target with 64 Gaussians, our feed-forward predictor ends 13% above the best per-image result we found after 16 internal steps and 5% after 32, on the same 100 holdout images; one seed, k-means width tuned on those images.”