While reaching outstanding prediction rates by means of black-box classifiers is relatively easy nowadays, reaching a proper trade-off between accuracy and interpretability might become a challenge. The most popular approaches reported in the literature to overcome this problem use post-hoc procedures to explain what the classifiers have learned. Less research is devoted to building classification models able to intrinsically explain their decision process. This paper presents a recurrent neural network---termed Evolving Long-term Cognitive Network---for pattern classification, which can be deemed interpretable to some extent. Moreover, a backpropagation learning algorithm to adjust the parameters attached to the model is presented. Numerical simulations using 35 datasets show that the proposed network performs well when compared with traditional black-box classifiers.
|Title of host publication
|Intelligent Systems and Applications
|Kohei Arai, Supriya Kapoor, Rahul Bhatia
|Springer International Publishing
|Number of pages
|Published - 2021