We study the dynamics of a Hebbian ICA algorithm extracting a sin- gle non-Gaussian component from a high-dimensional Gaussian back- ground. For both on-line and batch learning we find that a surprisingly large number of examples are required to avoid trapping in a sub-optimal state close to the initial conditions. To extract a skewed signal at least examples are required for -dimensional data and
exam- ples are required to extract a symmetrical signal with non-zero kurtosis.