Self-organized direction aware for regularized fuzzy neural networks

The fuzzy neural networks are efficient hybrid structures to perform tasks of regression, patterns classification and time series prediction. To define its architecture, some models use techniques that fuzzification of data that can divide the sample space in grid format through membership functions. The models that use such techniques achieve results with a high degree of accuracy in their activities, but their structures can vary greatly when the number of features of the problem is high, making of fuzzy neurons an exponential relationship between the number of inputs and the membership functions numbers used in the model of the input space. A multi-neuron structure can make the training and update of parameters damaging to the model's computational performance, making it impossible to work with problems of high dimensions or even with a high number of samples. To solve the problem of the creation of structures of hybrid models based on neural networks and fuzzy systems this paper proposes the use of a novel fully data-driven algorithm. This algorithm uses an extra cosine similarity-based directional component to work together with a traditional distance metric and nonparametric Empirical Data Analytics to data partitioning and forming data clouds in the first layer of the model. Another problem that exists in fuzzy neural network models is that some of their parameters are defined at random, so they challenging to interpret and can introduce casual situations that may impair model responses. In this paper we also propose the definition of bias and weights of the neurons of the first layer using the concepts of the wavelet transform, allowing the parameters of the neurons also to be directly related to the input data submitted to the model. In the second layer, the unineurons aggregate the neurons generated in the first layer and a regularization function is activated to determine the most significant unineurons. The weights used in the third layer, represented by an artificial neural network with an activation function of type ReLU, are generated using the concepts of the extreme learning machine. To verify the new training approach for fuzzy neural networks, tests with real and synthetic databases were performed for pattern classification, which led to the conclusion that the cloud-based approach and neuron weights generation based on the data frequency of training proves that the accuracy of the model is adequate to perform binary classification problems.