Better Latent Spaces for Better Autoencoders
This addresses a specific bottleneck in anomaly detection for high-energy physics, but appears incremental as it builds on existing autoencoder methods.
The paper tackled the structural problem of autoencoders only working in one direction for anomaly searches at the LHC by deriving classifiers from Gaussian mixture and Dirichlet latent spaces, with the Dirichlet setup solving the issue and improving performance and interpretability.
Autoencoders as tools behind anomaly searches at the LHC have the structural problem that they only work in one direction, extracting jets with higher complexity but not the other way around. To address this, we derive classifiers from the latent space of (variational) autoencoders, specifically in Gaussian mixture and Dirichlet latent spaces. In particular, the Dirichlet setup solves the problem and improves both the performance and the interpretability of the networks.