Out-of-distribution Detection and Generation using Soft Brownian Offset Sampling and Autoencoders

Deep neural networks often suffer from overconfidence which can be partly\nremedied by improved out-of-distribution detection. For this purpose, we\npropose a novel approach that allows for the generation of out-of-distribution\ndatasets based on a given in-distribution dataset. This new dataset can then be\nused to improve out-of-distribution detection for the given dataset and machine\nlearning task at hand. The samples in this dataset are with respect to the\nfeature space close to the in-distribution dataset and therefore realistic and\nplausible. Hence, this dataset can also be used to safeguard neural networks,\ni.e., to validate the generalization performance. Our approach first generates\nsuitable representations of an in-distribution dataset using an autoencoder and\nthen transforms them using our novel proposed Soft Brownian Offset method.\nAfter transformation, the decoder part of the autoencoder allows for the\ngeneration of these implicit out-of-distribution samples. This newly generated\ndataset then allows for mixing with other datasets and thus improved training\nof an out-of-distribution classifier, increasing its performance.\nExperimentally, we show that our approach is promising for time series using\nsynthetic data. Using our new method, we also show in a quantitative case study\nthat we can improve the out-of-distribution detection for the MNIST dataset.\nFinally, we provide another case study on the synthetic generation of\nout-of-distribution trajectories, which can be used to validate trajectory\nprediction algorithms for automated driving.\n