Müller et al. (2022) — ARXIV-2201.05989
Key takeaways
- An input encoding, not an architecture: coordinates index a multiresolution hash table of trainable feature vectors, optimised by ordinary gradient descent alongside a network small enough to be fast — training in seconds, rendering in tens of milliseconds
- Collisions are tolerated rather than resolved. A hash smaller than the coordinate space must collide; the multiresolution structure lets the network disambiguate them, “making for a simple architecture”, and gradient descent weights each entry toward whichever colliding coordinate carries more loss
- Table size is the single dial trading quality against performance, with no other structural change — which is what makes the encoding drop-in
- Explicitly task-independent, demonstrated across several neural graphics primitives rather than one
Standing in the anthology
Read — NOTE-042.
A source of SOTA-205, with LIT-108, LIT-086 and LIT-109 — the four papers whose independent convergence the #121 audit found. This note supplies the hash grid and the one-dial quality/cost trade.
The observation that leaves the domain is the collision argument: an
approximation whose errors are separated by other parts of the system does not
need a resolution mechanism. The record’s low-precision and approximation
practices reason about error magnitude; this is about error
distinguishability, which is a different question and one LIT-102’s
per-hop analysis also touches.