Logic-driven autoencoders

Autoencoders are computing architectures encountered in various schemes of deep learning and realizing an efficient way of representing data in a compact way by forming a set of features. In this study, a concept, architecture, and algorithmic developments of logic-driven autoencoders are presented. In such structures, encoding and the decoding processes realized at the consecutive layers of the autoencoder are completed with the aid of some fuzzy logic operators (namely, OR, AND, NOT operations) and the ensuing encoding and decoding processing is carried out with the aid of fuzzy logic processing. The optimization of the autoencoder is completed through a gradient-based learning. The transparent knowledge representation delivered by autoencoders is facilitated by the involvement of logic processing, which implies that the encoding mechanism comes with the generalization abilities delivered by OR neurons while the specialization mechanism is achieved by the AND-like neurons forming the decoding layer. A series of illustrative examples is also presented.