SciPost Submission Page
Anomalies, Representations, and Self-Supervision
by Barry M. Dillon, Luigi Favaro, Friedrich Feiden, Tanmoy Modak, Tilman Plehn
Submission summary
| Authors (as registered SciPost users): | Barry Dillon · Luigi Favaro · Tilman Plehn |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2301.04660v1 (pdf) |
| Code repository: | https://github.com/bmdillon/AnomalyCLR |
| Data repository: | https://mpp-hep.github.io/ADC2021/ |
| Date submitted: | 2023-02-07 11:24 |
| Submitted by: | Dillon, Barry |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approaches: | Computational, Phenomenological |
Abstract
We develop a self-supervised method for density-based anomaly detection using contrastive learning, and test it using event-level anomaly data from CMS ADC2021. The AnomalyCLR technique is data-driven and uses augmentations of the background data to mimic non-Standard-Model events in a model-agnostic way. It uses a permutation-invariant Transformer Encoder architecture to map the objects measured in a collider event to the representation space, where the data augmentations define a representation space which is sensitive to potential anomalous features. An AutoEncoder trained on background representations then computes anomaly scores for a variety of signals in the representation space. With AnomalyCLR we find significant improvements on performance metrics for all signals when compared to the raw data baseline.
Current status:
Reports on this Submission
Report
I appreciate your patience in receiving this feedback. In this paper, the authors delve into the application of contrastive learning for anomaly detection in the field of High-Energy Physics (HEP). The manuscript is well-crafted, and I have only a few minor comments to add:
1. In Section 2: Delphes utilizes FastJet in the backend. However, the authors haven't provided a citation for this tool; only the anti-kT algorithm is referenced.
2. In Section 3.2: It would be beneficial if the authors could provide a more detailed explanation of anomaly augmentations. The current text might be misconstrued as suggesting that these augmentations are model-dependent, influenced by the choice of augmentation technique. Clarity on this aspect would be helpful.
3. As a general comment, it would be valuable for the authors to discuss the potential reusability of their proposed method beyond simply distributing the learned model. The authors highlight the applicability of their approach to experimental data, which is of great significance. However, there is no mention of how this methodology can be made publicly available for future Beyond Standard Model (BSM) inference beyond the scope of the experimental collaboration.
- What kind of metadata should be made available to define the boundaries of the features?
- What are the critical nuances that future users should consider to avoid potential extrapolation of the defined model?
These considerations would contribute to the practicality and broader adoption of the AnomalyCLR technique in the HEP community.
Report
The authors present a framework for anomaly detection using contrastive learning. The paper is well written and provides a contribution to the ongoing ML efforts in particle physics. Therefore, I am happy to recommend publication in SciPost after the authors address the following point:
The presented data augmentations for anomaly modelling are not exhaustive. The anomaly signatures that they discuss would also be hard to miss in a more traditional approach to BSM discrimination. Are there issues of the framework with respect to scalability, would adding an order of magnitude of scenarios in parallel change the ability for classification. Also, do the authors believe that this procedure would be indeed perform superior in comparison to GANs pivoting background uncertainties?
Additional minor point:
three physical augmentations the data: -> to the data?