Learning from Extremely Ambiguous Examples

Oded Maron - Artificial Intelligence Lab, MIT

In traditional supervised learning, each training example is given a
label.  There are many learning problems where it is not possible to
give a label to every example.  One way of modeling this ambiguity is
called Multiple-Instance learning.  Every example is actually a
collection (bag) of instances.  If a bag is labeled negative, then all
the instances in it are negative.  However, if a bag is labeled
positive, then at least one of the instances in it is positive.  I
will discuss Diverse Density, an algorithm for learning from
Multiple-Instance examples.  In addition, I will show that this model
of ambiguity is extremely useful in applications such as drug
discovery, stock prediction, and image database retrieval.  Finally, I
will propose a general architecture for a system that learns with
ambiguous examples.