Artificial neural system or neuro-computing is used in many applications by simulating the characteristics of the human brain. Artificial Neural Networks deal with knowledge in a symbolic form (Ultsch, 1995). They can solve non-linear problems often bett er than conventional methods and are capable to approximate non-linear relations in data. In addition, incomplete and imprecise data can be processed. Artificial learning algorithms can be subdivided into two types, supervised and unsupervised. Supervised learning is the more useful technique when the data samples have known outcomes that the user wants to predict. On the other hand, unsupervised learning is more appropriate when the user does not know the data samples. Prior categorical division may not be clear because the problem may be a new one, for which the user has little experience. In such a case, an unsupervised learning procedure can provide insights into groupings that may make physical sense and facilitate future analysis. Neural networks ca n be applied to classification, association, perception of speech and vision, and reasoning. Neural networks can also be widely applied in expert systems and in signal processors such as aid for medical diagnosis.
Kohonen's Self-Organizing Feature Maps (SOFM) have the property that can be used to discover structures in high dimensional data and map them into a lower dimensional space. The SOFM algorithm was introduced by Professor Kohonen in 1981. Several major ap proaches to the contemporary artificial neural networks field, and new technologies have been based on SOFM algorithm. The most important practical applications are in exploratory data analysis, pattern recognition, speech analysis, robotics, industrial and medical diagnostics, instrumentation and control.
The basis in transforming sub-symbolic knowledge into symbolic knowledge is the distributed representation of the raw data in Artificial Neural Networks with collective behavior. Unsupervised learning neural networks can adapt to structures inherent in t he data. They exhibit the property to produce their structure during learning by integrating many case data. It cannot, however, look into the activity or weights of single neurons. Because of this users of the neural networks need tools to detect the str ucture in large neural networks. One way to do this could be thru graphical visualization tools. The methods called U-Matrix developed by Professor Ultsch in 1990 can be used to detect the structure of large two-dimensional Kohonen maps (Ultsch, 1985). Us ing the U-matrix method, a trained feature map is transformed into a landscape with "hills" or "walls" separating different regions where cases are located. All cases that lay in common "valley" or "basin" are considered to have a strong similarity such a s common structural properties.
It is becoming increasingly apparent that without some form of explanation capability, the full potential of trained artificial neural networks may not be realized. The problem is the inherent inability to explain in the comprehensible form, the process by which a given decision or output generated by an ANN has been reached. In symbolic AI systems, ANN has no explicit declarative knowledge representation. Therefore they have considerable difficulty in generating the required explanation structures. By b eing able to express the knowledge embedded within the trained artificial neural network as a set of symbolic rules, the rule-extraction process may significantly enhance the capabilities of ANNs to explore data to the benefit of the user.
Alternatively the system user may be able to use the extracted rules to identify regions in input space which are not presented sufficiently in the existing ANN training set data to supplement the data set accordingly. After learning the SOFM, an inducti ve extracted rule takes the training data, with the classification detected through SOFM, as an input, which generates rules for characterizing and differentiating the classes of the data. These rules can be interpreted by the expert of the domain or can be used by knowledge-based system.
Statement of the Problem
There is an increasing volume of data being generated and stored in computers. This can lead to the problem of being data rich but information poor. Hence a technique has to be devised to convert volume data into high-value information, which can be used to aid decision making in a database management process. The amount of information in database is growing at a rapid rate for which a knowledge extraction method is very essential. The database systems today offer very little functionality to support dat a mining applications. This may be the reason why massive amount of data are still largely unexplored and are either stored primarily in an off-line storage media. In the past, clustering large data sets has been a domain of classical statistical methods. More recently a new approach, Kohonen's Self-organizing Feature Map (SOFM) has been proposed in order to classify high-dimensional data sets, while having no analogous traditional method for unsupervised data analysis.
This work is aimed at demonstrating an approach on how artificial neural network technology can help to address the problem of converting the large volume of data into high value information. The study of knowledge extraction in unsupervised learning arc hitecture and rule extraction consists of three major modules:
Scope of the Study
The study will focus on the neural classifier of Kohonen self-organizing feature map.
The methodology is applied using existing land cover data of Cambodia.