🚀 go-pugleaf

           Department of Computer Science, University of Toronto
              (GB = Galbraith Building, 35 St. George Street)

       -------------------------------------------------------------

                                    AI
                  GB221, at 11:00 a.m., 11 February 1993

                              Radford M. Neal
           Department of Computer Science, University of Toronto

      " Learning Concepts with Hierarchical Latent Variable Models "

    To make sense of the world, we must learn how to describe the multitude
    of things we sense using concepts that capture the underlying
    structure that is present.  Often, we must do this in an unsupervised
    fashion, without any external guidance as to what concepts are
    appropriate.  Understanding and automating this process are major
    challenges for AI.

    In scientific investigations, exploratory data analysis has a similar
    aim, and has long been tackled using semi-automatic methods.  One
    approach used explains the relationships among a number of independent
    items by postulating unobservable "latent variables", associated with
    each item, that determine the distributions for the observed
    attributes.  The simplest such models employ a single, discrete "latent
    class" variable.  Many of the similarities between individual animals
    can be explained by the concept of a "species", for example.

    I will briefly describe work I have done on Bayesian latent class
    analysis, which includes a radical approach to the vexing problem of
    determining the appropriate number of latent classes.  I also describe
    work on learning more general latent variable models expressed using
    "belief networks", in which an item can be in more than one "class".
    To explain the symptoms of patients in terms of underlying diseases,
    for example, one must allow for the possibility that the patient has
    more than one disease.

    Mostly, however, I will discuss a hierarchical generalization of
    latent class models, in which similarities in the descriptions of
    the various classes are expressed in terms of classes at a higher
    level.  The practice in biology of grouping several species into
    a "genus", several genera into a "family", etc. is an example of
    such a model.  High-level classes can be defined in terms not only
    of concrete similarities, but also on the basis of more abstract
    characteristics, such as which attributes distinguish between the
    low-level classes, and which are irrelevant.  Consequently, rather
    than merely providing a coarser level of interpretation, such a
    high-level class structure can also bias the construction of low-level
    classes in favour of those with a coherent high-level structure.