SciPost Submission Page
Better Latent Spaces for Better Autoencoders
by Barry M. Dillon, Tilman Plehn, Christof Sauer, Peter Sorrenson
This Submission thread is now published as
|Authors (as registered SciPost users):
|Barry Dillon · Tilman Plehn
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.
Published as SciPost Phys. 11, 061 (2021)
Submission & Refereeing History
You are currently on this page
Reports on this Submission
- Cite as: Anonymous, Report on arXiv:2104.08291v1, delivered 2021-07-25, doi: 10.21468/SciPost.Report.3293
1. Well motivated, and excellent description of the motivation.
2. Topical, as illustrated by the other papers that were coordinated with this one.
3. Very clearly described methodology and ML structure.
4. Very nicely formatted, clear plots, a pleasure to read.
1. Maybe it would have been worthwhile to come up with a concrete signal that might have been missed by existing searches but discovered by a strategy taking advantage of these VAEs.
This is an excellent and very topical paper which will be of great value to the particle physics machine learning community, and I wholeheartedly recommend it for publication. I think there is a lot more to do be done in understanding how the structure of probabilistic latent spaces can be used for unsupervised event classification, and this paper is likely to inspire future work on the topic.