Enhanced Deep Type-2 Fuzzy Logic System For Global Interpretability

The recent advances in the field of Artificial Intelligence (AI) have led to the rapid deployment of AI systems in a variety of fields such as healthcare, financial, education etc. However, many of the AI systems are black boxes which restricts the use of these AI in applications that are highly regulated (such as financial, justice, medical, autonomous vehicles etc.) where it is necessary to provide satisfactory explanations for the decisions taken. A variety of approaches that have been proposed to tackle this problem, but these approaches generally emphasize providing satisfactory explanations for individual predictions at the cost of providing explanations at the global level. Hence, to solve these problems, in this paper, we present a hybrid deep learning type-2 fuzzy logic system which addresses these challenges by providing a highly interpretable model that can be trained using both labelled and unlabeled data. We also present a method to extract global and local explanations for this model. We also show that the presented model has reasonable performance when compared to stacked autoencoders deep neural networks.