FastICA

FastICA is an efficient and popular algorithm for independent component analysis invented by Aapo Hyvärinen at Helsinki University of Technology.^[1][2] Like most ICA algorithms, FastICA seeks an orthogonal rotation of prewhitened data, through a fixed-point iteration scheme, that maximizes a measure of non-Gaussianity of the rotated components. Non-gaussianity serves as a proxy for statistical independence, which is a very strong condition and requires infinite data to verify. FastICA can also be alternatively derived as an approximative Newton iteration.

Algorithm

Prewhitening the data

Let the $\mathbf{X} := (x_{ij}) \in \mathbb{R}^{N \times M}$ denote the input data matrix, $M$ the number of columns corresponding with the number of samples of mixed signals and $N$ the number of rows corresponding with the number of independent source signals. The input data matrix $\mathbf{X}$ must be prewhitened, or centered and whitened, before applying the FastICA algorithm to it.

• Centering the data entails demeaning each component of the input data $\mathbf{X}$, that is,

$x_{ij} \leftarrow x_{ij} - \frac{1}{M} \sum_{j^{\prime}} x_{ij^{\prime}}$

for each $i =1,\ldots,N$ and $j = 1, \ldots, M$. After centering, each row of $\mathbf{X}$ has an expected value of $0$.

• Whitening the data requires a linear transformation $\mathbf{L}: \mathbb{R}^{N \times M} \to \mathbb{R}^{N \times M}$ of the centered data so that the components of $\mathbf{L}(\mathbf{X})$ are uncorrelated and have variance one. More precisely, if $\mathbf{X}$ is a centered data matrix, the covariance of $\mathbf{L}_{\mathbf{x}} := \mathbf{L}(\mathbf{X})$ is the $(N \times N)$-dimensional identity matrix, that is,

$\mathrm{E}\left \{ \mathbf{L}_{\mathbf{x}} \mathbf{L}_{\mathbf{x}}^{T} \right \} = \mathbf{I}_N$

A common method for whitening is by performing a eigenvalue decomposition on the covariance matrix of the centered data $\mathbf{X}$, $E\left \{ \mathbf{X} \mathbf{X}^{T} \right \} = \mathbf{E}\mathbf{D}\mathbf{E}^T$, where $\mathbf{E}$ is the matrix of eigenvectors and $\mathbf{D}$ is the diagonal matrix of eigenvalues. The whitened data matrix is defined thus by

$\mathbf{X} \leftarrow \mathbf{E}\mathbf{D}^{-1/2}\mathbf{E}^T\mathbf{X}.$

Single component extraction

The iterative algorithm finds the direction for the weight vector $\mathbf{w} \in \mathbb{R}^M$ that maximizes a measure of non-Gaussianity of the projection $\mathbf{w}^T \mathbf{X}$, with $\mathbf{X} \in \mathbb{R}^{N \times M}$ denoting a prewhitened data matrix as described above. Note that $\mathbf{w}$ is a column vector. To measure non-Gaussianity, FastICA relies on a nonquadratic nonlinearity function $f(u)$, its first derivative $g(u)$, and its second derivative $g(u)^{\prime}$. Hyvärinen states that the functions

$f(u) = \log \cosh (u), \quad g(u) = \tanh (u), \quad \text{and} \quad {g}'(u) = 1-\tanh^2(u),$

are useful for general purposes, while

$f(u) = -e^{-u^2/2}, \quad g(u) = u e^{-u^2/2}, \quad \text{and} \quad {g}'(u) = (1-u^2) e^{-u^2/2}$

may be highly robust.^[1] The steps for extracting the weight vector $\mathbf{w}$ for single component in FastICA are the following:

1. Randomize the initial weight vector $\mathbf{w}$
2. Let $\mathbf{w}^+ \leftarrow E\left\{\mathbf{X} g(\mathbf{w}^T \mathbf{X})^T\right\} - E\left\{g'(\mathbf{w}^T \mathbf{X})\right\}\mathbf{w}$, where $E\left\{...\right\}$ means averaging over all column-vectors of matrix $\mathbf{X}$
3. Let $\mathbf{w} \leftarrow \mathbf{w}^+ / \|\mathbf{w}^+\|$
4. If not converged, go back to 2

Multiple component extraction

The single unit iterative algorithm estimates only one weight vector which extracts a single component. Estimating additional components that are mutually "independent" requires repeating the algorithm to obtain linearly independent projection vectors - note that the notion of independence here refers to maximizing non-Gaussianity in the estimated components. Hyvärinen provides several ways of extracting multiple components with the simplest being the following. Here, $\mathbf{1}$ is a column vector of 1's of dimension $M$.

Algorithm FastICA

Input: $C$ Number of desired components

Input: $\mathbf{X} \in \mathbb{R}^{N \times M}$ Prewhitened matrix, where each column represents an $N$-dimensional sample, where $C <= N$

Output: $\mathbf{W} \in \mathbb{R}^{C \times N}$ Un-mixing matrix where each column projects $\mathbf{X}$ onto independent component.

Output: $\mathbf{S} \in \mathbb{R}^{C \times M}$ Independent components matrix, with $M$ columns representing a sample with $C$ dimensions.

 for p in 1 to C:
    $\mathbf{w_p} \leftarrow$ Random vector of length N
    while $\mathbf{w_p}$ changes
        $\mathbf{w_p} \leftarrow \frac{1}{M}\mathbf{X} g(\mathbf{w_p}^T \mathbf{X})^T - \frac{1}{M}g'(\mathbf{w_p}^T\mathbf{X})\mathbf{1} \mathbf{w_p}$
        $\mathbf{w_p} \leftarrow \mathbf{w_p} - \sum_{j = 1}^{p-1} \mathbf{w_p}^T\mathbf{w_j}\mathbf{w_j}$
        $\mathbf{w_p} \leftarrow \frac{\mathbf{w_p}}{\|\mathbf{w_p}\|}$

 Output: $\mathbf{W} = \begin{bmatrix} \mathbf{w_1}, \dots, \mathbf{w_C} \end{bmatrix}$

 
 Output: $\mathbf{S} = \mathbf{W^T}\mathbf{X}$

See also

• Unsupervised learning
• Machine learning
• The IT++ library features a FastICA implementation in C++

References

1. Hyvärinen, A.; Oja, E. (2000). "Independent component analysis: Algorithms and applications" (PDF). Neural Networks 13 (4–5): 411–430. doi:10.1016/S0893-6080(00)00026-5. PMID 10946390.
2. Hyvarinen, A. (1999). "Fast and robust fixed-point algorithms for independent component analysis" (PDF). IEEE Transactions on Neural Networks 10 (3): 626–634. doi:10.1109/72.761722. PMID 18252563.

External links

• FastICA package for Matlab or Octave
• fastICA package in R programming language
• FastICA in Java on SourceForge
• FastICA in Java in RapidMiner.