Deep Differentiable Logic Gate Networks

Framework by Petersen et al. using logic gates as neurons with differentiable training, integrated into DiffLogic CA

Neighborhood — ranked by edge-count

thinker

Felix Petersen
introduces
Developer of Deep Differentiable Logic Gate Networks, whose framework is integrated in this work

concept

Continuous Relaxation of Logic Operations
implements
Key technique enabling gradient-based training of discrete logic gates by replacing binary operations with differentiable approximations
Probabilistic Gate Selection
implements
Each gate maintains a 16-dimensional probability distribution over binary operations, updated via gradient descent
Binary Logic Gates
implements
Fundamental discrete computation units used as neurons in DLGN and DiffLogic CA

framework

Differentiable Logic Cellular Automata (DiffLogic CA)
implements
The novel framework introduced in this paper, combining DLGN and NCA for fully differentiable discrete CA learning

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.

Differentiable Logic Gate Networks have not previously been shown to work in recurrent settings prior to this workclaim0.801
Authors' novelty assertion establishing the gap filled by DiffLogic CA
DiffLogic CA represents, to best of authors' knowledge, the first exploration of differentiable logic gate networks in a recurrent settingfinding0.772
Novelty claim about the contribution to the field
Different tasks yield distinct distributions of logic gate types between perception kernels and update networksclaim0.749
Interpretive claim based on circuit analysis across experiments
Logic Gate Perception Kernelsmethod0.737
Named technique in DiffLogic CA where fixed-structure logic circuits replace Sobel filters for neighborhood perception
Deep Belief Networkframework0.708
The integration of differentiable logic gates and neural cellular automata is a potential step towards programmable matter — Computroniumclaim0.708
Authors' broader vision claim linking their system to Toffoli and Margolus's programmable matter concept
Deep belief network (DBN)method0.706
Deep architecture with recurrent connections within layers, can learn compressed representations and retain stable attractors.
Deep Learningconcept0.706
Learning hierarchical representations of non-decomposable functions; proposed as formal equivalent to ETI process.