Biomedical Engineering Reference
In-Depth Information
The challenge of using a neural network to recognize and categorize data, especially novel data that
haven't been presented to the system before, is that of validating the results and of communicating
the rationale behind the results to the user. The greatest drawback of neural networks is that it's
practically impossible to assess the significance of what's happening inside of a complex network.
Even though the "wiring" of the nodes may be known, the relevance of changes in the strength of the
connections is difficult to assess, even when the strengths are known. As a result, the inner workings
of a neural network are difficult to validate.
Because a pure neural network presents such a formidable validation challenge, many neural network
data-mining systems are used in conjunction with rule-based expert systems that contain human-
readable rules in the form:
IF condition THEN outcome
These hybrid systems can categorize novel patterns and provide researchers with insight into the
operation of the biological system. The challenge in creating hybrid classification systems is
integrating the neural networks and rule-based expert systems in a way that doesn't compromise
classification performance while providing enough information on internal operation to allow the user
to assess the validity of the classification results. One approach to maximizing performance is to
develop a neural network and then use a tool that converts the network into a rule base that can be
compiled in C++ or Assembly language.
Statistical Methods
The statistical methods used to support data mining are generally some form of feature extraction,
classification, or clustering. Statistical feature extraction is concerned with recovering the defining
data attributes that may be obscured by imperfect measurement, improper data processing, or noise
in the data.
A variety of statistical pattern-classification methods may be applied to data mining. For example,
probabilistic classifiers are based on the principle that a pattern should be assigned to the class that
is most probable. Bayesian techniques that estimate the joint probability of distributions can also be
used to assess this probability. Although this method of classification generally provides excellent
results, it has a major drawback of requiring more complete data than other methods.
