Challenges for Unsupervised Anomaly Detection in Particle Physics

Abstract

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

What's Anomalous in LHC Jets?

“…In spite of this wealth of possible practical applications, the fundamental question still needs to be studied, namely what defines an anomaly search at the LHC? For large and stochastic datasets, the concept of outliers is not defined, because any LHC jet or event configuration will occur with a finite probability, especially after we include detector imperfections [51][52][53][54][55]. In this situation, a simple, working definition of anomalous data is an event which lies in a low-density phase space region.…”

“…A similar approach was discussed in Ref. [55], where instead of k-means clustering a k-medoids algorithm was used to obtain the representatives of the background dataset. The MinD score assumes that regular datapoints are close to the k-means centroids, whereas outliers are not.…”

Machine Learning in the Search for New Fundamental Physics

“…Since the initial proposals of using autoencoders for anomaly detection [40][41][42], a number of improvements and modifications have been suggested. An important observation is that autoencoders can be biased by the relative complexity of anomalous and background data, potentially leading to outliers with a lower loss than the background [129][130][131]. As the latent space in VAEs [33,34] is optimised to follow a known distribution for backgrounds, it can also be used as anomaly score [132,133].…”

"Flux+Mutability": A Conditional Generative Approach to One-Class Classification and Anomaly Detection

“…Examples of OCC can be found in an extensive literature review provided by [1]. As for AD, there is a growing number of applications that span from accelerator operations to physics analyses, the latter being of great interest for example at the Large Hadron Collider (LHC) since new physics beyond Standard Model (BSM) remains elusive (as discussed, e.g., in [2][3][4][5]). 1 In both cases, one typically deals with multiple features that vary as a function of the phase space of the final state particles reconstructed in the detector.…”