Learning Concepts through Multi-Class Diverse Density
View/ Open
Date
2008Author
Hiransoog, Chalita
Metadata
Abstract
This research investigates the possibility of creating an intelligent system based on the
philosophy that the world is ambiguous and a system gains knowledge by learning from
these ambiguous examples where the learning can especially be improved when a system
is allowed to play an active role in requesting these ambiguous examples. The above
philosophy will bridge the gap between the traditional Artificial Intelligence (knowledgebased
AI) and the behaviour-oriented Artificial Intelligence (intelligence emerging from
behaviour). Concept learning, due to its simplicity and features needed to prove this philosophy,
is chosen as the studied platform. Based on the aforementioned philosophy, the
task of concept learning is comparable to the multiple-instance learning framework where
the learning framework will be modified to tackle more classes compared the the original
two-class problem, named here as the multi-class problem. The multi-class multipleinstance
learning problem is thus defined. One of the methods used to solve the original
multiple-instance learning framework, the Diverse Density method, is selected due to its
simplicity, robustness, and incremental property. The method is then modified to solve the
newly defined multi-class multiple-instance learning problem. To explore the functionality
and the efficiency, the modified method, multi-class Diverse Density, was tested on
both artificial data and real-world applications: stock prediction task, assembly task, and
document search. It was found that redefining the two-class problem as multi-class problems
allows a wider range of ambiguous concepts to be better captured than is possible
with the original multiple-instance learning framework. Moreover interactivity, the ability
to play an active role in requesting or suggesting examples to learn, was proven to enhance
the learning process when integrated into the multi-class Diverse Density method.
In summary this research proves that the task of concept learning of ambiguous objects
can be solved using the proposed multi-class Diverse Density method where the added
interactivity feature improves the learning further