We propose a novel method for the injection of symbolic knowledge into neural networks allowing data scientists to control what the network should (not) learn. Unlike other similar approaches, our method does not (i) require ground input formulæ, (ii) impose any constraint on the NN undergoing injection, (iii) affect the loss function of the NN. Instead, it acts directly at the backpropagation level, by increasing the penalty whenever the NN output is violating the injected knowledge. Experiments are reported to demonstrate the potential (and limits) of our approach.
evento origine
