Speaker
Description
Data is what drives all research forward, regardless of scientific discipline. At the Large Hadron Collider (LHC) in Switzerland, the data stream of registered events can reach about 60 million megabytes per second, making it physically impossible to save all the produced data with current storage technology. This means that data selection has to be performed at an early stage in the experimental process, often using so-called trigger systems. While these are sophisticated, interesting data may be omitted by accident. This could potentially mean losing out on new discoveries, and thus a risk of not reaching the goal of the LHC in itself.
Data compression techniques can reduce the size of data drastically while giving a sufficiently faithful representation of the uncompressed data. For high-energy physics, using new data-compression techniques as a part of the data selection process would allow for further storage savings without waiting for major technological advancements in storage media, and thus allowing for new scientific discoveries to be made earlier by increasing the amount of data that can be recorded. One of the compression techniques that have been under recent investigation uses autoencoder networks. This is a machine-learning based approach, which utilizes a certain type of neural network known as an autoencoder.
Autoencoders are, in their most basic form, a neural network with multiple layers where the number of inputs is equal to the number of outputs, and the input and target datasets are the same. If the dimension of the hidden layer is (much) smaller than the dimension of the input and output layers, an autoencoder will be tasked with finding an effective representation of the input data, which can then be reconstructed in the output layer. As such, autoencoders are a good candidate for data compression.
In this contribution, autoencoder-based compression will be evaluated, by using the results presented in the thesis "Deep Autoencoders for Compression in High Energy Physics" (https://lup.lub.lu.se/student-papers/search/publication/9004751) by Eric Wulff as a foundation. In addition, we will also present new results and give an outlook in terms of autoencoder compression in the context of future LHC and new colliders.
