Feature Representations Using the Reflected Rectified Linear Unit (RReLU) Activation

Deep Neural Networks (DNNs) have become the tool of choice for machine learning practitioners today. One important aspect of designing a neural network is the choice of the activation function to be used at the neurons of the different layers. In this work, we introduce a four-output activation function called the Reflected Rectified Linear Unit (RReLU) activation which considers both a feature and its negation during computation. Our activation function is "sparse", in that only two of the four possible outputs are active at a given time. We test our activation function on the standard MNIST and CIFAR-10 datasets, which are classification problems, as well as on a novel Computational Fluid Dynamics (CFD) dataset which is posed as a regression problem. On the baseline network for the MNIST dataset, having two hidden layers, our activation function improves the validation accuracy from 0.09 to 0.97 compared to the well-known ReLU activation. For the CIFAR-10 dataset, we use a deep baseline network that achieves 0.78 validation accuracy with 20 epochs but overfits the data. Using the RReLU activation, we can achieve the same accuracy without overfitting the data. For the CFD dataset, we show that the RReLU activation can reduce the number of epochs from 100 (using ReLU) to 10 while obtaining the same levels of performance.