Out-of-distribution detection methods on deep neural network encodings of images and tabular data

The scope of this thesis work is to implement an anomaly detector pipeline, capable of detecting the anomalous and normal samples in the test set of a deep learning classifier. The pipeline is divided into two steps: the preprocessing part, the idea in this case is to use the compact representation given by the inner layer of a trained neural network on a "clean" dataset, and an anomaly detector, which scope is to recognize if the samples is an anomaly. The work was initially developed during the internship experience in the start-up Axyon AI. The application of the algorithm in this context is the recognition of economic anomalies. The idea is to use the pipeline created paired with a classifier to give an insight to the "quality" of the data where the classifier make the prediction.