Challenges for unsupervised anomaly detection in particle physics

Anomaly detection relies on designing a score to determine whether particular event is uncharacteristic of given background distribution. One way define use autoencoders, which rely the ability reconstruct certain types data (background) but not others (signals). In this paper, we study some challenges associated with variational such as dependence hyperparameters and metric used, in context anomalous signal (top $W$) jets QCD background. We find that hyperparameter choices strongly affect network performance optimal parameters for one are non-optimal another. exploring networks, uncover connection between latent space autoencoder trained using mean-squared-error transport distances within dataset. then show representative events dataset can be used directly anomaly detection, comparable autoencoders. Whether autoencoders or best represent necessarily identification. These unsupervised bolster case additional exploration semi-supervised alternative approaches.