Neural networks, linear functions and neglected nonlinearity (Q2386624)

The article focuses on the ability of feedforward networks to approximate a linear function. By means of MAPLE empirical experiments one justifies the following statement: neural networks can represent linear functions without the need for skip-layer connections. In case skip-layer formulation is used, there are potential problems in separating linear and nonlinear components. Statistical tests with neural networks make use of these issues.