State-of-the-art comparison, experiments, and future work on symbolic knowledge extraction for neural networks

Modern learning models and machine learning suffer from a significant issue called opacity ("black box"), which prevents you from employing them in sensitive situations like legal or medical. Lack of formality and expressive ability is "black-box" algorithms' primary flaw, which restricts their range of applications. It's possible that the rules provided by NN rule extraction methods are easier to comprehend than the rules provided by decision trees. In addition to maintaining the high accuracy of the networks, using NNs to build the rules may result in the rules presenting new perspectives on the data. With the help of neural networks, this thesis aims to define what logical rule extraction is, describe the types of rules that may be extracted, and examine various extraction method approaches. For the purpose of extracting rules from a neural network with a single hidden layer, two algorithms—one for classification (Neurolinear) and the other for regression (REFANN)—are also presented. You will be shown the experiments done, the outcomes, and comparisons with other extraction algorithms. These experiments are carried out using PSyKE. It is designed as a library that provides a general-purpose API for knowledge extraction along with a variety of algorithms that implement it in order to support classification and regression issues. The difficulties encountered when applying the pruning method for REFANN and the application of the discretization algorithm of the activation function for Neurolinear are noted in the conclusion. The top data extractors for classification are CART and CReEPY, whereas the best extractor for regression is CREEPY, according to a comparison of the various extraction approaches. CReEPY has been found to be the top extractor on BB for both classification and regression. Neurolinear was shown to be the worst extractor for classification on both data and BB. This comparison was carried out using the Qs score measure.