Machine learning is now a well developed field that enables complex data sets to be analyzed with a new effectiveness. This research is concerned with further developing the fundamental algorithmic components of machine learning so as to extend the reach of these methods to new applications, particularly those of a cognitive nature where many complex tasks need to be learned simultaneously from a limited data set. Emphasis will be placed on techniques that allow learning from a few examples, and those that allow learning of multi-object relations. The mistake-bounded model of machine learning has provided some powerful techniques for showing that certain knowledge representations, such as linear threshold functions, can be learned attribute-efficiently. The task remains, however, to classify more precisely the known representation classes according to this property. On the subject of learning relations, a self-evident challenge is to control the computational complexity of this task. However, this is subsumed by the more general question of how exactly one should define this problem in order that the algorithmic techniques of machine learning can be brought to bear effectively on it.
