مقاله انگلیسی رایگان در مورد شبکه عصبی احتمالاتی وزن دار – الزویر 2018

Classical feedforward neural networks such as multilayer perceptron or radial basis function network have their layers linked using weighted connections. Within the training process, the utilized weights must be first initialized and then iteratively recomputed to optimize some assumed performance measure of the model for a given training data. The probabilistic neural network [36,37], however, is the model unequipped with any additional weighting factors inside its structure. PNN is therefore free from time consuming weights’ update. This fact gives this network the advantage in application popularity. The usage of PNN can be found in the domains of medical diagnosis and prediction [9,15,17,18,20], image classification and recognition [4,25,45], earthquake magnitude prediction [1], multiple partial discharge sources classification [43], interval information processing [12–14], phoneme recognition [7], email security enhancement [41], intrusion detection systems [40], classification in a time-varying environment [27,28] or hardware implementation [46]. The variant of PNN in regression tasks, known as general regression neural network [38], is also studied by many authors, e.g.: in function approximation [10] or knowledge discovery in data streams [6]. Four layers create a structure of a conventional PNN: the input layer represented by data attributes, the pattern layer composed of as many neurons as training patterns, the summation layer containing a single neuron for each class and the output layer with a decision neuron which determines the classification outcome. In the research centered on PNN architecture, some minor emphasis is usually placed on applying the weights inside this network. The first contribution related to PNN’s weights is presented in [22]. However, the coefficients are not computed explicitly. The network operates using anisotropic Gaussians to provide the output for a considered sample by utilizing the covariance matrix instead of a single smoothing parameter in the activation function. In the works of [34] and [35], the weighting coefficients are directly introduced inside the PNN model. They are placed between pattern and summation layer. These factors are calculated from soft labeling probability matrix based on Bayesian algorithm. In turn, the authors of [23–25] create weighted PNN based on the class separability in the data. The weight coefficients are defined as the ratio of ‘between-class variance’ and ‘withinclass variance’ for a particular training pattern. They are included between a pattern and summation layer of PNN.