Summary · Chapter 5

Is this a neural representation?

Read-out: F̂(x) = Σk rk(x) zk, rk = mkφk / Σj mjφj; on the grid F̂ = AZ. Code: FineRefineModel.forward, free_gaussian_fine.py. Size at 64 Gaussians: 24,960 numbers against 393,216 in the 32×32 grid.

Dense MSE: mean over tokens and channels of (F̂ − T)² on holdout images. Targets have unit-length rows, so ×384 gives squared error per token as a fraction of its squared length.

64 GaussiansTargetDense MSE×384
Gridless64, least squares16×160.00010414.0%
Current model16×160.00005872.3%
Current model32×320.00011374.4%

800 holdout images, one seed, holdout reused; different targets are not comparable.

Question (Chapter 1)Answer
How is it read?Fixed normalised-mixture rule over explicit Gaussians; defined at any x (accuracy between tokens not yet measured, roadmap P3)
How are parameters found?Placement: amortised network pass. Features: closed-form ridge solve against the teacher, per image

Terminology: hybrid representation in the neural-fields survey's terms. “Semi-implicit” clashes with semi-implicit variational inference (Yin and Zhou, 2018): define it, or say “hybrid explicit–implicit feature field”.

Caveat: the teacher is needed at inference (placement reads its tokens; the solve needs them). Roadmap P4.

Defence answer: “It is a hybrid: continuous like a neural field, but made of explicit, interpretable Gaussians read by a fixed rule. Placement is amortised; features are the closed-form optimum given placement, so today the teacher is needed at inference. The contribution is placement; teacher-free features are the next step.”

Source: Xie et al., Neural Fields in Visual Computing and Beyond (2022).