Laboratory of Computer and Information Science / Neural Networks Research Centre CIS Lab Helsinki University of Technology

Icasso documentation


About documentation:
These web pages only give a general view to Icasso. The detailed documentation of the functions is found in the help texts in MATLAB, e.g.,
  >> help icassoStruct

Introduction

Icasso is based on running FastICA several times (resampling). Icasso pools all the estimates together and forms clusters bottom-up among them. The basic idea is that a tight cluster of estimates is considered to be a candidate for including a "good" estimate. A centroid of such cluster is considered a more reliable estimates than any estimate from an arbitrary run. (Instead of an average as a centroid, Icasso visualizes and returns a centrotype from each cluster. This is the one of the original estimates that is most similar to other estimates in the same cluster. You can compute the average by using Icasso functions.)

The basic procedure

Icasso is a sequential procedure that is split into several phases (functions). In general, Icasso consists of the following steps:
  1. Parameters for the estimation algorithm(s) are selected: e.g., for FastICA the estimation approach (symmetrical or deflatory), contrast function, etc. The estimation is run N times using the selected training parameters. Each time the data is bootstrapped and/or the initial conditions of the estimation algorithm are changed.
  2. Mutual similarities between all the estimates are computed. As the measure of similarity, we use the absolute value of the linear correlation coefficient between the independent components. The estimates are clustered according to their mutual (dis)similarities. In principle, the clustering method can be freely selected. We apply agglomerative clustering with average-linkage criterion.
  3. The clustering is visualized as a dendrogram and a 2D plot. The user investigates how dense the clusters are. The clustering of the estimates is expected to yield information on the reliability (robustness) of estimation. A compact cluster emerges when a similar estimate repeatedly comes up despite of the randomization.
  4. The user can retrieve the estimates belonging to certain cluster(s) for further analysis and visualization.
Read illustrative examples on using Icasso in the publications.

Some parameters

Firstly, you have to select the parameters for FastICA. In particular,
  1. the (reduced) data dimension (d) that may be less than the original input data dimension (PCA dimension reduction is often applied in FastICA) and
  2. the number of ICA estimates (m) extracted on each resampling cycle.
are of interest here.

For Icasso you have to select also

  1. the resampling mode,
  2. the number of resampling cycles (N), and
  3. the number of estimate-clusters (L).

Resampling mode

Yon can use
  1. both a different random initial condition and resampling of the data (by bootstrapping) in each resampling cycle,
  2. different random initial condition for FastICA on each resampling cycle but keep the training data set fixed, or
  3. fixed initial condition in each cycle but bootstraps every time the data.

Number of resampling cycles (N)

Basically, the more cycles the better. However, Icasso uses currently hierarchical clustering which causes a computational bottleneck. Icasso can currently handle a moderate total number of estimates M, say, 1000-2000, and consequently, a moderate number of resampling cycles (N). For example, if you extract 15 independent components at one resampling cycle 50 resamplings might be appropriate M=mN=15x50=750.

Number of ICA estimates (estimate-clusters) (L)

Often, ICA is performed so that the number of the components is the same as the input data dimension (possibly after PCA dimension reduction) m=d. If you use L=d=m it means that you try to find as many estimates as there are data dimensions - and the quality index and centroid estimate for all of these. The default in Icasso is to set the number of estimate-clusters L=d.

In FastICA, you can extract less independent components than there are dimesnsions in FastICA (m < d). In Icasso, you can also freely select the number of estimate-clusters. For example, you can run FastICA in the deflatory mode and extract, e.g., only one component at each run but extract several "robust" estimates after Icasso. You can also group the estimates to bigger or smaller number of estimate-clusters. Interpreting the results is up to you.


Results

Sources, demixing matrix (W), and mixing matrix (A)

FastICA estimates the demixing matrix (W). In the Icasso procedure this is done several times, and the estimates are clustered. Icasso returns a centroid (centrotype) estimate W from each estimate-cluster. This should represent a more reliable estimate than any single estimate from one run of FastICA. You can also return all estimates in a cluster by using appropriate Icasso functions.

However, the computational results that Icasso give do usually not represent a strictly orthogonal base in the whitened data space since they are directly the natural centroids (centrotypes) of the estimate-clusters. You have to orthogonalize the result in an appropriate manner if necessary.

The mixing matrix A is a pseudoinverse of W and the sources are returned by computing S=WX by using the original data that is stored in Icasso data structure.

Estimate stability index (Iq)

Icasso returns a stability (quality) index (Iq) for each estimate-cluster. This gives a rank for the corresponding ICA estimate. In the ideal case of m one-dimensional independent components, the estimates are concentrated in m compact and close-to-orthogonal clusters. In this case the index to all estimate-clusters is (very close) to one. The value drops when the clusters grow wider and mix up.

R-index

R-index should be addressed only in exploratory work (if wish to explore different clustering solutions). The R-index is a heuristic Davies-Bouldin type relative measure for a "natural" number of clusters.

Local minima of this index are "good" solutions in terms of having mutually isolated "natural" clusters.

As any relative clustering validity index The index is heuristic and should be used only as a guideline. If the structure of the estimate space is complex, this index is dubious.


Implementating the procedure using Icasso functions

See script megdemo for example.

First step (icassoEst) is to compute randomized ICA estimates N times from data X using function icassoEst. Output of this function (we will use variable name sR is called Icasso result data structure. It logs all the methods and parameters used in the process, and the results from the Icasso procedure. You can extract information from this data structure either directly or by using functions icassoResult and icassoGet.

The batch of Icasso functions that perform similarity computation, clustering and the 2D projection are collected in icassoExp

Finally, you can explore the clustering and get the results by launching icassoShow. You can examine rel ationships between estimates and clusters in detail.

Functions that start with string icasso are main functions: they use the Icasso result structure as input and/or output.

icassoEst
FastICA parameters and resampling
[icassoStruct]
This is a subfunction automatically called by icassoEst. However, its help text describes in detail the Icasso data structure with reference to the Icasso process. This might be of interest to you if you are going extract information directly from the data structure.
icassoExp
Performs clustering and projections for visualization
icassoShow
Visualizes and returns results
icassoResult
Returns results
Function icasso implements the basic procedure from resampling to visualization in one batch.
More functions and details of the Icasso process

You are at: CIS → Documentation

Page maintained by webmaster@cis.hut.fi, last updated Tuesday, 21-Dec-2010 15:55:12 EET