Many Gaussians: summing, compositing and sharing
| Rule | Value at x | Behaviour |
|---|---|---|
| Summation | Σk ck gk(x) | Fades in gaps, piles up in overlaps |
| Alpha compositing | Σk ck αk Πj<k(1 − αj) | Needs a front-to-back order; fades where nothing is opaque |
| Normalised sharing (this project) | Σk rk(x) ck | No gaps, no double counting, no order |
Responsibility: rk(x) = mkφk(x) / Σj mjφj(x) = softmax over k of log mk + log φk(x). In a Gaussian mixture p(x) = Σ mkφk(x) it is the posterior probability that component k produced x (Bayes' rule).
Soft partition: each Gaussian owns a soft region. Narrow widths give sharp, nearest-centre boundaries; the widest Gaussian wins far away; larger mass pushes boundaries out. Example: Gaussians at −0.5 (width 0.2) and 0.4 (width 0.3), equal masses, split at −0.11.
The matrix A (P tokens × K Gaussians, rows sum to one; 1024 × 64 on the 32×32 grid): for each cell, average mkφk over a 2×2 pattern (±¼ cell), then normalise. Code: patch_quadrature, patch_log_density, gaussian_responsibilities in geometry.py; FineGrid.points() in free_gaussian_fine.py.
Defence answer: “The read-out is a normalised mixture: every location is a convex combination of Gaussian features, weighted by mixture-model responsibilities. It needs no depth order, leaves no gaps, and makes each Gaussian own a soft region of the image.”
Source: Bishop, Pattern Recognition and Machine Learning (2006), section 9.2.