Is AI helping research or compromising it?

The Large Hadron Collider (LHC). Source: CERN.

When artificial intelligence (AI) helps with research, does AI’s limits compromise it? The question emerged earlier this month in Science - The Wire. One can’t help but wonder. Is AI really improving research, and if it does, how? The best way to assess machine learning’s impact is to dive at the heart of algorithms with an expert.

Jennifer Ngadiuba is a Caltech postdoctoral fellow working on the Compact Muon Solenoid (CMS) experiment, an international collaboration involving 5000 scientists and engineers at CERN, the world’s largest particle physics laboratory in Geneva.

The CMS is one of the detectors at the Large Hadron Collider (LHC), a huge scientific instrument that accelerates and smashes particles together to explore the mysteries of the universe. The CMS detector, a magnet that can bend the path of particles, helps scientists gather clues.

Ngadiuba is part of a group of physicists pushing AI to the edge. Physics is a way for her to satisfy her thrive for understanding. She guides us through the high-energy particle accelerator and searching for the unknown in data.

Geneva Solutions: What was your first encounter with AI?

Jennifer Ngadiuba: Probably when I realized that Facebook could recognize all faces!  I saw the potential. But I really started working with it at CERN. I realized we had just started to scratch the surface of AI in physics.

GS: How does AI come into play at the LHC?

JN: The LHC collides protons. Every time there is a collision, the energy is transformed into matter. To identify the particles, we need high-resolution detectors and sensors, as each collision produces thousands of particles, we need really high resolution detectors. The main goal is to answer fundamental questions about the universe. This can be achieved by directly finding new particles. One of our latest and highest achievements was the discovery of the Higgs boson. It completed the Standard Model, the model we have of particle physics since the ’50s, which is very well consolidated at this point. The Higgs boson was the missing piece of the puzzle.

The problem is that there are many phenomena in nature that we don’t understand, like dark matter or dark energy, and the fact is that the matter as we know it (protons, neutrons, electrons,…) constitutes only a small fraction of the universe, the rest is a mystery. So what we would expect from this collision is to create a dark matter particle, and to observe it. It happens but it happens very rarely. It’s a big challenge because to be sure we see it, when it happens, we need to collect a lot of data. And it’s big data. Just to give you an idea, we now collect hundreds of petabytes of data every year. By 2026, as the collider becomes more powerful, we will be talking about exabytes.

So it becomes more challenging in terms of data. That’s where AI comes into play.

GS: How does it happen?

JN: At the moment, we produce 1000 particles per event (collision). With a more powerful accelerator, we’ll produce five times more particles. So you can imagine how many particles will come out of it. In order to understand them, we need detectors with a higher resolution. The data will become more and more high dimensional, more and more complicated. Of course, you can engineer physics-inspired algorithms to make something out of it but it will take ages. Artificial Intelligence is a shortcut.

We are also examining ways to apply AI to simulations, typically when we simulate our physics data. It takes a lot of computing time, and the computing time depends on how much data you want to reproduce. One shortcut is to use generative adversarial models. These models can be used to generate your event. It’s much faster. Once you train the algorithm, the AI, you’re done. You pass your data through it, it is very fast. It is parallelizable.

We are trying to replace many traditional algorithms with highly parallel networks, also because we can now exploit the hardware. AI hardware is improving and developing very fast. We now have very powerful GPUs, and are also exploring FPGAs, ASICs, and exotic hardware architectures… Since there is a lot of development in the industry we can exploit their development in terms of AI software for the hardware that we have in mind. We don’t have to rewrite traditional algorithms in the specific hardware language.

GS: The human brain has unlimited cognitive capacities. As the AI is a shortcut, does it compromise the results?

JN: You can learn from what the AI has learnt. It’s not a black box. But, for us, the gain is related to our goal, which is to find new particles. If we don’t apply AI to take these shortcuts, we might lose the new particle, or not see it at all. What’s fundamental is to observe new phenomena, from that we learn. You can also go back, after having used all your shortcuts, and check how the algorithms worked to infer the properties of the new phenomena or particle.

There are massive amounts of theories that can predict these new particles. We don’t know which theories are true. By depending on one theory, you tailor your algorithm for this or for that. But if you use unsupervised machinery algorithms, you train them to learn the background, the standard model, what you already know. Once you pass a new physics event, it will be easier to find an anomaly. And it is the anomaly we search for, something that is not what we know.

This approach is new for us, and I think, well I hope it is very promising. Finding something we haven’t thought of yet.

GS: What’s the biggest challenge with AI?

JN: The industry developments are so fast. Sometimes I feel that we apply something that was thought of 10 years ago and we still don’t know how to apply it to our problems. We have to adapt very quickly. We keep following what comes out from the industry, try to apply, see if it works for us. We have several contacts within the industry and try to invite experts for talks and tutorials. And we do the same for the industry. Some of our projects are so cutting edge that the industry doesn’t always think of certain developments.

GS: Can you understand that AI scares people?

JN: My mom is one of them! Of course, it’s a powerful tool, it has a lot of potential to help human life, society. But as any powerful tool in the wrong hands it can have disruptive consequences. So one of our priorities at CERN is that all collaborations we set up are for the good, not to damage anybody. For example helping a car company develop safer cars, exploiting the power of AI for preventing disasters, or understand climate change. You can’t just say no to AI. Not employing its benefits for our advantage would be a bold move given its potential to make the world a better place.