Feed-forward neural networks: a geometrical perspective (Q4713640)

The classical question posed by Radon is the following: Consider \(n\) points \(P_i\subset R^d\) in general position (\(n\) points in \(R^d\) are in general position if no \(d+1\) in a \((d - 1)\)-plane, no \(d\) in a \((d - 2)\)-plane, and so on) and calculate the number of different partitions of \(\{P_i\}\) induced by hyperplanes. The answer obtained by many authors [see \textit{L. Schläfli}, Theorie der vielfachen Kontinuität (Bern, 1852); reprinted in Gesammelte mathematische Abhandlungen. Bd. I. Basel: Birkhäuser Verlag (1949; Zbl 0035.21902)] was that this number equals \[ \sum_{k=0}^ d \binom{n-1}{k} \tag{1} \] [the author of this paper refers to \textit{Th. M. Cover}, IEEE Trans. Electron. Comput. 14, 326--334 (1965; Zbl 0152.18206)]. Result (1) is not valid for the set of vertices of a \(d\)-dimensional hypercube \((d>2)\). The author shows that for \(n<O((d+1)^{3/2})\) the number of linear separations of \(n\) random hypercube vertices tends to that of \(n\) points in general position (i.e. (1)).