finding

active

finding:checkerboard-circuit-trained-on-16x16-grid-successfully-generalizes-to-64x64-grid-with-4x-more-time-steps

Checkerboard circuit trained on 16x16 grid successfully generalizes to 64x64 grid with 4x more time steps

Grid scaling generalization result demonstrating boundary-size invariance

Source paper

extracted_from

Differentiable Logic Cellular Automata: From Game of Life to pattern generation with learned recurrent circuits

Neighborhood — ranked by edge-count

Concepts (1)

concept

Boundary-Size Invariance
supports
Property where a rule learned on fixed-size grid generalizes to larger grids, observed in checkerboard and lizard experiments

Questions (1)

question

Could the discretization and minimal circuit size be finding some minimal procedural description for generating patterns of interest?
supports
Open question raised by the 5-gate checkerboard solution and its generalization properties

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Learned checkerboard generation circuit reduces to just 5 active logic gates after pruning (6 with one redundant AND)finding0.811
Remarkably minimal circuit discovered for checkerboard pattern generation
Asynchronous DiffLogic CA requires 50 steps vs 20 steps for synchronous training to reconstruct checkerboard patternfinding0.780
Cost of asynchronous training in terms of convergence time steps
DiffLogic CA with 128-bit cell state and 12 steps successfully learns 20x20 lizard pattern, generalizing to 40x40 gridfinding0.765
Demonstration of DiffLogic CA on complex non-regular shapes with arbitrary memorization requirements
Checkerboard pattern formation exhibits directional propagation from bottom-left to top-right despite no built-in directional bias in the modelfinding0.764
Emergent property observed in checkerboard pattern generation
Alternation of squares and rectangles was superior to simple checkerboard for Martinez floorfinding0.746
In pattern trials, alternating square and rectangular elements felt more harmonious than a plain checkerboard.
Total active gates in checkerboard circuit is 22 (excluding pass-through gates A and B)finding0.741
Circuit size finding for checkerboard experiment before pruning
DiffLogic CA with 64 channels successfully learns 16x16 colored Google G pattern in 15 steps using 8-color binary palettefinding0.735
Demonstration of multi-channel RGB color pattern generation with binary states
DiffLogic CA fully converges to 16x16 checkerboard target pattern with both soft and hard losses reaching zerofinding0.734
Core result of pattern generation experiment demonstrating recurrent circuit learning