Abstract
Deep neural networks have shown great success in various industry sectors such as security and healthcare sectors. Despite their widespread adoption for biometric user authentication sys tems and clinical applications, most deep neural networks have limited interpretability and are not robust to outlier samples during deployment. Therefore for critical applications involving high-stakes decision making, many domain experts still prefer shallow machine learning meth ods that are generally more intuitive and interpretable than deep neural networks. We address the issue of transparency in prediction and the detection of outlier test samples when building a deep learning application system. We first approach the problem by investigating the application of deep neural networks in mouse dynamics-based biometric authentication systems and generate explanations to explain the contribution of the pixels to the prediction when mouse trajectories are converted into im ages. This enables us to verify what features are learned by the model from the training sam ples. Subsequently, we study the behaviour of unsupervised outlier detection models on more challenging outlier definitions beyond outliers defined by semantic classes to understand the behavior of the outlier models before introducing a model towards joint example-based expla nation and outlier detection. We stabilize a deep one-class classification method by proposing regularizers and observe substantial performance gain over the standard variant. Ultimately, we propose a scalable and unified interpretable deep neural network that is capable of explanation and detecting out-of-distribution samples for an arbitrary prediction task. We address the issues of prediction transparency and the unreliable model outputs caused by anomalous samples by demonstrating the practicality and relevance of our proposed network for an application involv ing the binary classification of the presence of metastatic tissue in breast cancer