“…The width of the neural network can be tradedoff by depth to alleviate the worst-case requirement for the number of neurons (Anthony 2010). Other neural networks have also been proven to universally implement boolean formulas, such as the binary pi-sigma network (Shin and Ghosh 1991), the binary product-unit network (Zhang, Yang, and Wu 2011) A body of empirical work followed the positive theoretical results on the learnability of boolean functions: (Miller 1999) shows that parity and multiplier functions are efficiently learnable, (Franco and Anthony 2004;Franco 2006;Franco and Anthony 2006) study complexity metrics that related to the the generalisation abilities of boolean functions implemented via neural networks, (Subirats et al 2006;Subirats, Jerez, and Franco 2008) and (Zhang, Ma, and Yang 2003) show algorithms for learning boolean circuits with thresholding neural networks, while (Prasad and Beg 2009) studies pre-processing techniques for using ANNs to learn boolean circuits and in (Beg, Prasad, and Beg 2008) they study approximating a boolean function's complexity using an ANN. (Pan and Srikumar 2016) showcases how neural networks with ReLU activation implement boolean functions much more compactly than with threshold linear units.…”