Classification characteristics of SOM and ART2

and therefore adapts the resulting network to these patterns. Artificial neural network algorithms were originally designed to model human neural activities. They attempt to recreate the processes involved in such activities as learning, short term memory, and long term memory. Two widely used unsupervised artificial neural network algorithms are the Self-Organizing Map (SOM) and Adaptive Resonance Theory (ART2). Each was designed to simulate a particular biological neural activity. Both can be used as unsupervised data classifiers. This paper compares performance characteristics of two unsupervised artificial neural network architectures; the SOM and the ART2 networks. The primary factors analyzed were classification accuracy, sensitivity to data noise, and sensitivity of the algorithm control parameters. Guidelines are developed for algorithm selection. Introduction The training of artificial neural networks can be separated into two main categories: supervised and unsupervised. Supervised training requires prior knowledge of what the network is expected to do. The network must be trained to map the exemplars of the training set to known outcomes. Supervised network training can monitor the convergence of the network toward the expected outcome and use it as a criterion to stop training. Unfortunately, because the training is focused on the expected outcome, unexpected possibilities may be either incorrectly mapped or are rejected as noise. The prior knowledge associated with the training set creates a classilication bias in the network. Unsupervised training requires no prior knowledge of the problem domain. The network groups exemplars in the data set with other exemplars having similar characteristics. Competitive training is the procedure normally used to control training in unsupervised machine learning algorithms. In competitive training, the output node with the "best" or maximal output is selected for training. Other nodes may receive reduced training or no training at all. This training algorithm reinforces dominant patterns in the training data Permission to copy wi~out fee all or par o f this material is granted provided that the copies are not made or distributed for direct commercial advantage, the ACM copyright notice and the title of the publication and its date appear, and notice is given that copying is by permission o f the Association for Computing Machinery. To copy otherwise, or to republish, requkcs a fee and/or specific permission. O 1994 ACM 089791,-647.,.6/94I 0003 $3.50 Because prior knowledge is not used in the training of the network, bias toward specific expectations is not introduced. This does not mean that no bias exists. A critical assumption is that the training data set is representative of the problem domain. When the training data is not totally representative of the problem domain, the network's classification accuracy is affected. Also, since there is no expected outcome, stopping criterion other than the convergence of the network restJlts must be chosen. Typically, the algorithm is trained for a predetermined number of epochs. This criterion can sometimes allow the network to overtrain. This paper focuses on comparing the performance characteristics of two unsupervised artificial neural network architectures; the Self-Organizing Map (SOM) and the Adaptive Resonance Theory (ART2) networks. The SOM maps vectors from an n-dimensional space to a twodimensional output network. The ART2 algorithm maps vectors from an n-dimensional space to a one-dimensional" output network. The primary factors analyzed were classification accuracy, sensitivity to data noise, and sensitivity of the algorithm control parameters. Accuracy was determined by. comparing algorithm classifications to the expected classifications. Self-Organizing Map Many neurological systems exhibit a self-organizing inclination. Nerves in the human auditory system are organized so that neighboring nodes respond to similar sound frequencies. This spatial distribution inspired Teuvo Kohonen to develop his Self-Organizing Map [6]. The SOM is a competitive unsupervised learning algorithm. Mutual lateral interactions are developed between the output nodes by training neighborhoods of nodes to respond to an input vector. The size of these training neighborhoods linearly decreases over a training session. The result is a trained network where neighboring nodes share similar properties and distant nodes are obviously different. This property is what allows SOM to organize and group input data over its two-dimensional output surface. The SOM defines a map