Hard Sciences Being Shaken by Machine Learning

Particle physicists have taught algorithms to solve previously unsolvable issues and take on whole new challenges

From email to the Internet, particle physicists have historically been early users of technology, if not its creators. Therefore, it is not unexpected that researchers began training computer models to identify particles in the chaotic jets produced by collisions as early as 1997. Since then, these models have labored along and have become better and better, albeit not everyone has been happy with this advancement.

Particle physicists have taught algorithms to solve previously unsolvable issues and take on whole new challenges during the past ten years, along with the larger deep-learning revolution.

The LHC needs to store 600 petabytes of data throughout the upcoming few years of data collecting, even with an effective trigger. Therefore, researchers are looking for ways to condense the data.

To begin with, the data utilized in particle physics differs greatly from the conventional data used in machine learning. Though convolutional neural networks (CNNs) have excelled in categorizing photos of commonplace items like trees, kittens, and food, they're less good at handling particle collisions. Javier Duarte claims that the issue is that collision data from sources like the Large Hadron Collider isn't, by nature, an image.

Flashy representations of LHC collisions may deceitfully fill the whole detector. In actuality, a white screen with a few black pixels represents the millions of inputs that aren't actually registering a signal. Although the resulting image is substandard because of the weakly provided input, a more recent design known as graph neural networks can make good use of it.

Innovation is needed to overcome more particle physics problems. According to Daniel Whiteson, "We're not merely importing hammers to smash our nails." We need to create new hammers since there are strange new types of nails. One odd aspect is the large amount of data produced at the LHC—roughly one petabyte per second. Only a small portion of this vast volume of high-quality data is retained. To create a better trigger system that keeps as much outstanding data as possible while removing low-quality data, researchers are working to educate a sharp-eyed algorithm to sort better than one that is hard coded.

However, according to Duarte, such a program would need to execute in just a few microseconds to be efficient. Particle physicists are pushing the boundaries of machine methods like pruning and quantization to accelerate their algorithms to solve these issues. Researchers are looking for ways to compress the data since the LHC needs to store 600 petabytes during the next five years of data collecting (equal to around 660,000 movies at 4K resolution.

Particle physicists can now approach the data they utilize differently, thanks to machine learning. They are learning to take into account the numerous other events that take place during a collision rather than concentrating on a single event, such as a Higgs boson degrading to two photons. Researchers like Thaler are now adopting a more comprehensive view of the facts, rather than merely the fragmented point of view that results from evaluating occurrences interaction by interaction, even though there is no causal link between any two events.

There is already a stigma surrounding coding, which is occasionally dismissed as "not real physics," and similar worries abound over machine learning. One concern is that machine learning may muddle physics, making analysis a black box of automated operations that are difficult for humans to comprehend.