CAN I MAKE MY NEURAL AND NEURO-FUZZY SYSTEMS A BIT MORE USEFUL?

摘要

Neural networks are easy-to-use wonderful tools for many applications. However, design of any useful "intelligent system" using neural networks raises several important issues. This talk will begin with a brief discussion on these issues with emphasis on three of them, (ⅰ) An important system design principle is: "Make everything as simple as possible, but not simpler." (Einstein). This suggests to use a simple architecture. This can be achieved by discarding poor features, reducing the use of dependent (redundant) features, and pruning nodes of the network, as much as possible, (ⅱ) The network should refuse to make any decision when faced with unfamiliar data. In other words, the network should not make a decision when it is given a test data point that is far from the training data that were used to design the network, (ⅲ) The network should be capable of incremental learning. For example, in an industrial application, a new defect may appear and users should be able to augment the network to classify this new defect.