Heuristic self-organization of knowledge representation and development: analysis in the context of explainable artificial intelligence

Abstract

From the analysis of the main theoretical provisions of heuristic self-organization systems and logical models, it follows that according to O. G. Ivakhnenko's systems of heuristic self-organization, the first task is to determine the factors content “that determine the essence of different images”. These are the images that characterize the objects of a particular subject area. After determining the composition and content of these images, the next problem is solved, namely, the problem of “generating the new successful heuristic”, which in content is a solution that leads to increased accuracy. Note that we are talking about improving the accuracy of solving the problem of data processing. It follows from the above mentioned that heuristic self-organization systems are data processing systems. This allows the multiplicity of heuristics. Heuristics in content correspond to the logical rules applied in heuristic self-organization systems. The main provisions of the heuristic self-organization system theory were developed by O. G. Ivakhnenko in the eighties of the last century, but they remain unnoticed to this day. At this time, the task is to explain why the neural network makes such a decision and not another. Based on this, the concept of “explainability of artificial intelligence” was introduced for artificial intelligence. It is the content of heuristics that forms the structure of the neural network in the form of logical rules and determines the logic of the decision made. It is established that the derivation rule, which is the basis for constructing artificial neural networks, is an abductive rule, which, unfortunately, does not meet the fourth heuristic and does not meet the definition of intelligence: intelligence is the ability to measure things. Unfortunately, none of the neural networks can measure things. From the analysis of the basic rules content of inference, it follows that the dialectical method of inference is general (generating) for the basic logical methods of inference. The difference lies in the composition and content of the middle member of the triangular relationship, namely, in the form of the element combination of the relationship: the transition from one concept to another. The explainability of artificial intelligence refers to the laws of the structure and activity of artificial neural networks. But modern theories of artificial neural networks ignore the existence of logical rules (heuristics), which were established by O. G. Ivakhnenko. After all, only knowing the rules based on which problems are solved, it is possible to check the correctness of the decision, but not by searching for such rules. The three hypotheses about the explainability of artificial intelligence and the theory of machine identification can be further defined as statements or theorems and strictly proved.