Prediction Using a Fuzzy Inference System in the Classification Layer of a Convolutional Neural Network Replacing the Softmax Function

In this work, a new fuzzy inference system type-1 is presented which predicts each of the classes to which each weight belongs, in this way the replacement of the Softmax activation function that it used in the classification layer, which is responsible for predicting the percentage of membership of each of the last weights of the network within the classification layer of a convolutional neural network. The neural network has been trained with different epochs from 10 to 60 training epochs, showing results not as similar as when using the classical Softmax function inside the classifier layer of the network. This network has a depth of 2 convolution layers, 2 pooling layers and 1 classification layer, in this last layer is where the proposed fuzzy inference system is implemented for the replacement of the Softmax that is in charge of predicting a percentage which will pass to the classification. Applied to a sample of 3 classes from the ORL database.