Towards a Framework for Interpretation of CNN Results with ANFIS

The performance of Deep Convolutional Neural Network (CNN) in the field of computer vision and image processing has shown tremendously amazing results. Nevertheless, the internal hidden and deep network architecture of CNN offers a Black Box computational model. Thus, it is infeasible to interpret and explain the computed results directly from such a model, whilst semantic interpretation and explanation are often required for applications such as fault monitoring in a physical plant and medical diagnosis in a human body.
To address this challenging problem, various approaches have been proposed in the emerging area of eXplainable AI.
Adaptive Network-based Fuzzy Inference System (ANFIS) provides a possible modelling technique that returns transparent and interpretable reasoning results, but it tends to work with data of a dimensionality considerably lower than that workable with CNN. Yet, an integration of these two different approaches may present a meaningful solution to the problem. This paper introduces such a novel framework which enables an interpretable explanation of the CNN outcomes with an ANFIS. This is facilitated by the use of Fuzzy Rough Feature Selection that converts high dimensional data into low dimensional data while minimising information loss. Initial experimental results illustrate the working of this framework.