T. Magoč (USA)
Intelligent systems, neural networks, fuzzy integration, dependence
Neural networks have found applications in numerous areas in real life including pattern recognition, portfolio optimization, and weather forecast. Even though neural networks have had great success in some of the applications, they are not without drawbacks. One of the main drawbacks is that they are prone to overfitting to training data samples, and thus might not perform well when applied to real life examples. One of the reasons for overfitting to training samples is that neural networks aggregate input values by using weighted sum, which assumes that all inputs are independent. Thus, while a neural network tries to “learn” weights to assign to each input, the dependencies among inputs are neglected and thus the weights are modified to fit to training samples rather than to adequately represent a given real life situation. To solve the issue of neglecting the dependencies among the input parameters, we propose the use of fuzzy integration instead of weighted sum as the aggregation operator at the first level of nodes. Fuzzy integration takes into consideration dependencies among inputs and therefore is expected to reduce overfitting to training samples.
Important Links:
Go Back
