The paper in [4] is absolutely fascinating - I'm very much a neophyte here, but I believe it shows that practically any ReLU network can be represented as a tropical ratio of two tropical polynomials, and thus can be analyzed with geometric principles including visualizations of surfaces. It's been cited in more recent work: https://scholar.google.com/scholar?cites=1003719112553620451... - does anyone know if there's been any significant advancements here?