Relational Reasoning for Out Of Distribution detection without fine-tuning

Despite the remarkable results achieved across multiple fields, modern Deep Neural Networks can still prove to be extremely untrustworthy when operating with data coming from a different distribution compared to the one they were trained on. Such a limitation constitutes a major issue in safety-critical applications, where the reliability of a model is a particularly important factor. The Out Of Distribution (OOD) detection task tackles the aforementioned problem by recognizing whether test-time data samples come from a known or an unknown distribution. In order to obtain acceptable performance, traditional methods usually require a training or fine-tuning stage of the network on the known data, but this may be impractical or even infeasible in many deployment scenarios (e.g. because of data privacy or computational constraints). This thesis focuses on the semantic aspect of OOD detection (i.e., identifying data samples belonging to unknown categories) and, more specifically, it studies how to avoid any new learning effort on the in-distribution data by exploiting pre-trained models based on relational reasoning. Rather than focusing on standard image classification, a relational reasoning model is trained by comparing pairs of samples. This task can be formulated in various ways depending on the choice of the exact optimization objective. In this work, several loss functions are considered and their impact is analyzed by studying the behaviour of the obtained models. Moreover, other competing approaches are presented in order to draw a comparison and help better delineate the current landscape for OOD detection methods.