I stumbled upon a great blog post which discusses t-SNE.  If you’re not familiar with it, t-SNE is an algorithm that take high dimensional data (like fMRI images) and projects them down to a 2D space which is easier to view.  Of course, to make that sort of projection, some information from the high dimensional space must be discarded.  The t-SNE algorithm is a technique for automatically deciding what information should be preserved.  


The blog post has a few sample datasets and allows you to perform t-SNE on those datasets while changing various parameter settings.  t-SNE cannot see the groupings (clusters) of the points (indicated by color).  You can judge a t-SNE projection by the separation of the clusters in the new 2D space.


What was particularly interesting to me was the situations where t-SNE “found” structure in random data.  For example, when it breaks clusters up into several smaller groups even though there should be no distinction within cluster.  This is important to note because we might try and draw conclusions from these sub-clusters when no relevant distinction exists.


t-SNE plots are interesting and useful for data exploration, but they should not be the only way you analyze your data, and the interpretations should always be taken with a grain of salt.  For further analysis, you might consider this toolbox for comparing dimensionality reduction techniques, including t-SNE.