notes-qualitativeUnsupervisedLearning

By analogy to 'qualitative physics', qualitative unsupervised learning discovers regularities in data, groups instances into a classification, and creates ontologies, but only when there is some evidence for a discrete separation between classes.

For example, in neuroscience, we may have a bunch of cells and we may be wondering if there are classes of subtypes. If we look at the distribution of the soma size and see a strongly multipolar distribution, that indicates discrete classes. But even if the distribution of soma size is unipolar, and the distribution of axonal length is also unipolar, perhaps if we plot both of these against one another, we might find a scatterplot showing two very lines. This then is also sufficient to conclude that there are different subtypes.

Another example is the 'single dissociation' and 'double dissociation' criteria for determining separation and independence, respectively, of functions.