Ephemeral Learning - Augmenting Triggers with Online-Trained Normalizing Flows

Abstract

The large data rates at the LHC require an online trigger system to select relevant collisions. Rather than compressing individual events, we propose to compress an entire data set at once. We use a normalizing flow as a deep generative model to learn the probability density of the data online. The events are then represented by the generative neural network and can be inspected offline for anomalies or used for other analysis purposes. We demonstrate our new approach for a toy model and a correlation-enhanced bump hunt.

Machine learning and LHC event generation

“…Finally, there are many simulation-related questions in fundamental physics, where AImethods allow us to make significant progress. Examples going beyond immediate applications to event generation include symbolic regression [131], sample and data compression [58,132], detection of symmetries [133][134][135][136], and many other fascinating new ideas and concepts.…”

MadNIS - Neural multi-channel importance sampling

“…It implies that a potentially expensive integrand f (x) has to be evaluated for every event used to train the network, which makes it inefficient. One way to alleviate this problem is to buffer already generated samples and use them for a limited number of training passes [18].…”

“…(1) using the trained weights defined in Eq. (18) we first define normalized channel-wise probability distributions as…”

“…Generally, it is possible to improve numerical integration through neural networks by directly learning the primitive function [10], or using modified and enhanced implementations of importance sampling [11][12][13][14][15][16]. Technically, this promising approach encodes a change of integration variables in a normalizing flow [17] and then uses online training [18] while generating weighted phase space configurations, or weighted events. This rough online training is successful because normalizing flows, or invertible networks (INNs) [19,20], are especially well-suited, stable, and precise in LHC physics applications [21].…”

Normalizing Flows for LHC Theory

“…Generative networks open new possibilities to enhance the efficiency of simulations [1,6]. They are able to learn underlying distributions with high precision [7,8,9,10,11,12,13,14] and can therefore provide more efficient phase space mapping [15,16,17,18,19,20], amplify [21,22] and compress data [23], serve as surrogate models in phenomenological studies and provide fast detector simulations [24,25,26]. Finally generative networks enable the inversion of the simulation chain [27,28,29].…”