The expressive power of a ReLU neural network is often measured by the number of linear regions into which it partitions the input space. A new paper on arXiv, announced as a replace-cross revision, takes up the question of how hard it is to count those regions computationally.

The abstract notes that many different, non-equivalent definitions of what counts as a linear region are in use. This variety suggests that the complexity of counting may depend heavily on the chosen definition, and the paper appears to investigate that relationship.

Because the full results are not detailed in the abstract, the specific findings remain unclear. The announcement signals a revised version of the work, but the core focus is on the computational complexity of a fundamental expressivity measure in deep learning theory.