06 - Fast ICA

How long did you take for the last exercise sheet

• I didn't do it.

• < 5 hours

• 5-8 hours

• < 8 hours

Scaling independent sources will make them less independent. (T/F)

• True.

• False.

Infomax: Why would we choose a sigmoid as our transfer function?

• To make the data less gaussian

• To decorrelate the data

• To approximate the cdf of a peaky pdf.

• To approximate the cdf of a gaussian.

Kurtosis can be used to measure ...

• uniformity.

• non-gaussianity.

• entropy.

• mutual information.

One disadvantage of kurtosis is ...

• division by zero.

• it's only applicable to Gaussians.

• sensitivity to outliers.

• computationally expensive.

Which of the following is "super-Gaussian"?

• Uniform

• Laplacian

• Poisson

• Gaussian with variance > 1.

FastICA: How do I get multiple sources instead of only one?

• Whiten the data.

• Repeat FastICA Algorithm for a different b (w).

• Through orthogonalisation.

• Subtract the first component and repeat.

FastICA: The unmixing matrix W sorts the independent components according to ...

• variance

• index

• there's no sorting

• lexical

FastICA: Which contrast function would you use for a data set with few outliers?

• All.

• None, give me a fourth.