Analyzing differentiable fuzzy logic operators (Q2060729)

The main aim of this paper is to propose an approach that integrates machine learning with symbolically expressed backgrond knowledge in the form of formulae of fuzzy logic. To this purpose, the authors analyze a large class of fuzzy logic operators (conjunctions, disjunctions, implications, and aggregations) from the point of view of their differentiability. The importance of this aspect becomes clear if we consider machine learning scenarios where, e.g., partial derivatives equal to zero on large portions of domains of these operators mean a loss of potentially useful information. The authors introduce \emph{Differentiable Fuzzy Logics} with semantics using vector embeddings. This means that objects in the domain are vectors of reals. Predicates are interpreted as functions that map these vectors to truth values from the interval \([0, 1]\). Based on this semantics, the fuzzy maximum satisfiability problem is considered. If all operators are differentiable, this problem can be solved using gradient descent. Particular attention is paid to fuzzy implications, since a significant proportion of background knowledge is expressed using implications. The product S-implication (known also as Rechenbach implication) \(I_\mathrm{RC}(a, c) = 1 - a + a\cdot c\) possesses favorable derivatives; however, there are several problems, in particular the behavior around \(a = 0 = c\). The authors tackle this problem by introducing a new class of implication operators called \emph{sigmoidal implications}. These are transformations of other fuzzy implications using the sigmoid function in such a way that the boundary conditions hold. The paper also contains a description of a series of experiments addressing a problem of semi-supervised learning using the MNIST dataset of handwritten digits. The results of these experiments can provide guidelines for choosing fuzzy logic operators in machine learning scenarios. The best results have been obtained for operators abandoning the normal symmetric configurations of these operators.