Neural networks for supervised learning and prediction, with applications to character recognition and medical database analysis.

摘要

ARTMAP is a class of neural network architectures that perform incremental supervised learning of recognition categories and multidimensional maps in response to input vectors presented in arbitrary order. The fuzzy ARTMAP neural network achieves a synthesis of fuzzy logic and adaptive resonance theory (ART) by exploiting a formal similarity between the computations of fuzzy subsethood and ART category choice, resonance, and learning. On a benchmark task of recognition of distorted letters of various fonts, fuzzy ARTMAP performance is compared to that of genetic algorithm systems. Fuzzy ARTMAP has an error rate that is consistently less than one third that of the best performing genetic algorithm classifiers.; A more difficult handwritten digit recognition task requires the identification of real, noisy digits from ZIP codes. Two preprocessing algorithms, based on positional or orientational information extracted from the image, are tested. The orientation selective algorithm proves to be more successful. Showing high recognition accuracy, fuzzy ARTMAP performance is improved by a modified learning rule, which enables the "forgetting" of insignificant information. The K-Nearest Neighbor (KNN) classifier evaluated in the study outperforms fuzzy ARTMAP but requires more memory and recognition time.; The last part of the dissertation introduces the ARTMAP-PI (Probabilistic Inference) neural network, which extends the capabilities of fuzzy ARTMAP to provide probabilistic outcome estimates. Two features distinguish ARTMAP-PI from fuzzy ARTMAP: distributed activity during performance, and a new layer in the architecture. The activity of the new instance-counting layer diminishes the influence of statistically less-significant information on probabilistic predictions made by the network. Also, the introduction of these new features allows for the incorporation of inconsistent data in the learning process. Fuzzy ARTMAP, ART-EMAP, and ARTMAP-PI performance is evaluated on databases that (a) differentiate malignant and benign tumors based on fine needle aspiration, (b) diagnose heart disease for cardiac patients, (c) predict the occurrence of complications after cholecystectomy, and (d) determine whether a patient will develop diabetes. The performance of the KNN classifier and the logistic regression model is evaluated and compared to that of ARTMAP-based classifiers. ARTMAP-based classifiers demonstrate similar or better performance compared to these other approaches.