Cancelout: A Layer For Feature Selection In Deep Neural Networks

Feature ranking (FR) and feature selection (FS) are crucial steps in data preprocessing; they can be used to avoid the curse of dimensionality problem, reduce training time, and enhance the performance of a machine learning model. In this paper, we propose a new layer for deep neural networks - CancelOut, which can be utilized for FR and FS tasks, for supervised and unsupervised learning. Empirical results show that the proposed method can find feature subsets that are superior to traditional feature analysis techniques. Furthermore, the layer is easy to use and requires adding only a few additional lines of code to a deep learning training loop. We implemented the proposed method using the PyTorch framework and published it online (The code is available at: www.github.com/unnir/CancelOut).