Improving out-of-distribution detection in deep models by latent space analysis

Despite their widely known success in various applications, deep models have vulnerabilities due to their inner mechanics. Sensitivities against adversarial and out-of-distribution (OOD) samples are some of the issues causing state-of-the-art networks to make wrong predictions with high confidence. These problems are critical for high stakes applications such as self-driving vehicles and medical imaging. The underlying reasons for this sensitivity could be explained by overfitting and the trade-off between generalization and robustness. Some studies investigate the latent spaces of the networks to gain a more in-depth understanding of why networks capture some details that lead to poor generalization. This thesis focuses on understanding the inner dynamics of deep neural networks (DNNs) by investigating their intermediate layer weights and activations. A latent space exploration method is designed to analyze the intermediate layer activations of DNNs. The adversarial robustness of DNNs is studied based on the observations from this analysis. Finally, an out-of-distribution (OOD) sample detection approach based on the intermediate layer activations and OOD samples is proposed. The proposed OOD detection method trains a set of prototype vectors on the in-distribution (ID) dataset. The prototype vectors are used to extract features from ID and OOD sample activations to train a multi-layer perceptron for discriminating ID and OOD instances. Experiments on several benchmark datasets show that the proposed OOD detection approach could perform state-of-the-art.