Universal approximation theorem for interval neural networks (Q1276145)

A 4-layer interval neural network is defined as interval function \[ \begin{multlined} {\mathbf y}= {\mathbf f}_{NN}(x_1,\dots, x_n)= [\underline f_{NN}(x_1,\dots, x_n), \overline f_{NN}(x_1,\dots, x_n)]=\\ \sum^L_{l= 1}{\mathbf W}_l\cdot f_0\Biggl( \sum^K_{k= 1} {\mathbf w}_{lk}'\cdot{\mathbf y}_k+{\mathbf w}_{l0}'\Biggr)+ {\mathbf W}_0,\end{multlined}\tag{1} \] where \({\mathbf y}_k= f_0 \left(\sum^n_{i=1} {\mathbf w}_{ki}\cdot x_i+{\mathbf w}_{k0}\right)\). Here the bold faced letters stand for real compact intervals, and all operations are interval operations. The function \(f_0\) is usually either a so-alled logistic function \(f_0(x)= 1/(1+ \exp(-x))\) or the identity \(f_0(x)= x\). The input signals \(x_1,\dots, x_n\) and the output \({\mathbf y}\) are given and the weights \({\mathbf w}_{ki}\), \({\mathbf w}_{lk}'\), \({\mathbf W}_l\) are wanted such that (1) holds at least approximately if \((x_1,\dots, x_n)\) varies over some given set \(M\). The authors show that for every continuous interval valued function \({\mathbf f}(x_1,\dots, x_n)= [\underline f(x_1,\dots, x_n),\overline f(x_1,\dots, x_n)]\) on a compact set \(M\subset \mathbb{R}^n\) and for every \(\varepsilon>0\) there exist weights for which for every \((x_1,\dots, x_n)\in M\) the inequalities \(|\underline f(x_1,\dots, x_n)- \underline f_{NN}(x_1,\dots, x_n)|\leq \varepsilon\) and \(|\overline f(x_1,\dots, x_n)- \overline f_{NN}(x_1,\dots, x_n)|\leq \varepsilon\) hold. They thus generalize the universal approximation theorem for the standard (i.e., non-interval) 4-layer neural network.