Multivariate approaches used in physics analyses by the High Energy Physics community often combine high-level observables estimated by very complex algorithms. The process to select these variables is usually based on a ``brute force'' approach, where all available event features are tested for multiple combinations of the algorithm hyperparameters.

In this work, we propose an original method based on the use of a CancelOut layer to select to give as input to a Fully Connected Neural Network. Promising results are obtained in the development of a DNN classifier to select proton-proton collisions where a boosted Higgs boson decay to two $b$-quarks.