Neural Network Computations with DOMINATION Functions

Abstract

We study a new representation of neural networks based on DOMINATION functions. Specifically, we show that a threshold function can be computed by its variables connected via an unweighted bipartite graph to a universal gate computing a DOMINATION function. The DOMINATION function consists of fixed weights that are ascending powers of 2. We derive circuit-size upper and lower bounds for circuits with small weights that compute DOMINATION functions. Interestingly, the circuit-size bounds are dependent on the sparsity of the bipartite graph. In particular, functions with sparsity 1 (like the EQUALITY function) can be implemented by small-size constant-weight circuits.