Introduction - Multimedia Database Retrieval: Technology and Applications - page 4

However, human beings are complex creatures, and their motivations and
behaviors are difficult to measure and characterize. As a result, the interpretation
criteria utilized by human users are not fixed. Human beings are capable of inter-
preting and understanding visual contents—to simultaneously synthesize context,
form and content—which is beyond the capability of any current computer method.
Human interpretation depends on individual subjectivity and the information needed
at a particular time and for a particular event. In addition, users learn from the
available information (search results) to recognize their needs and refine their visual
information requests (refine the queries) [ 8 ]. In other words, the interpretation of
visual content by a human user is non-stationary, or fuzzy, and is very difficult to
describe with fixed rules. A human user is an adaptive-learning component in the
decision-making system.
In order to build a computer system to simulate and understand the deci-
sion making processes of human beings, the above-mentioned characteristics of
adaptability should be taken into account. Learning to adapt and to optimize
decision-making is primary to the goal of creating a better computer-based retrieval
system.

1.3.1.1 User-Controlled Relevance Feedback

Chapter 1 of this topic, therefore, explores the development of an adaptive machine
that can learn from its environment, from both user advice as well as self-adaptation.
Specifically, the adaptive machine requires two important properties to achieve
this purpose: nonlinear decision-making ability, and the ability to learn from
different sources of information (i.e., multi-modeling recognition). By embedding
these two properties into the computer, the system can potentially learn what
humans regard as significant. Through a human-computer interactive process, the
system will develop the ability to mimic non-stationary human decision-making in
visual-seeking environments. The relevant topics include: Content-based similarity
measurement, using linear functions and nonlinear functions, Relevance feedback
(RF), Linear/non-linear kernel-based adaptive retrieval, Single-class Radial basis
function (RBF) network, RBF networks with adaptive learning, gradient-descent
learning, fuzzy-RBF with soft decision, and a Bayesian framework for fusion
of short-term relevance feedback (content information) and long-term relevance
feedback (context information).

1.3.1.2 Machine-Controlled Relevance Feedback

Chapter 2 introduces the automation process to optimize the learning system by
incorporating self-organizing adaptation into relevance feedback. This process is
referred to as pseudo relevance feedback (RF). Optimization is the process of
reducing the user's direct input, as well as adjusting the learning system architecture
for flexibility in practical use in multimedia retrieval. While user interaction