Boundary Optimizing Network (BON)

Despite all the success that deep neural networks have seen in classifying certain datasets, the challenge of finding optimal solutions that generalize well still remains. In this paper, we propose the Boundary Optimizing Network (BON), a new approach to generalization for deep neural networks when used for supervised learning. Given a classification network, we propose to use a collaborative generative network that produces new synthetic data points in the form of perturbations of original data points. In this way, we create a data support around each original data point which prevents decision boundaries to pass too close to the original data points, i.e. prevents overfitting. To prevent catastrophic forgetting during training, we propose to use a variation of Memory Aware Synapses to optimize the generative networks. On the Iris dataset, we show that the BON algorithm creates better decision boundaries when compared to a network regularized by the popular dropout scheme.