Artificial data was generated with a randomly initialised MLP network having one hidden layer with tanh-nonlinearities and a linear output layer. The network had a 2-10-5 structure and the data thus consisted of a two-dimensional manifold nonlinearly wrapped in five dimensions. The inputs had Gaussian distributions with unit covariance matrix. The five-dimensional outputs of the random network were further linearly embedded in a 10-dimensional space and finally, Gaussian noise was added to the data which consisted of 1000 vectors.