Barry Barish: AI “World Models” and Scientific Discovery in Physics and Astronomy

Okay, so first it looks to me like you're in 201 East Bridge, which is the room I grew up in at Caltech. I'm not sure because I can't see the whole room. I'm sorry I'm not there today. I happen to be in New York at this point and I must say I should acknowledge something which is very central here, which is today is 9-11, which is a day that we're acknowledging what happened in New York City. A couple of decades ago. Anyway, today I'm going to give you a short talk on kind of two topics that relate to fundamental science and applying AI in general to applications that are like scientific applications, I think. One is the starting development of what's called the world model for AI, and I'll give you an example of what we are doing now and what needs to be done. Second, I'm going to address the question of how in the realm of using AI we can confidently make scientific discoveries and know that they're reliable. I'm going to address those both as unsolved problems, but problems that are being worked on. First, when we talk about what a world model is, I just summarize it quickly. As an AI world model is an internal representation of an environment that you're trying to model that enables AI to predict how it will change as a consequence of any possible action. That's the goal of a world model. The world itself can be quite limited. It could be a detector, which is the example I'll give you in a minute, and the model need not be complete or explicitly encode all the physical laws. For example, we don't have to teach it Newton's laws of gravity or something in order to use what is the world model. The world model is a pretty confined thing that describes the local world that we're going to deal with to solve a certain problem. I'm going to take as an example a forefront effort presently beginning to be underway. The role of AI or a world model so far are in the planning in developing this facility. Everyone knows there's particle accelerators for both particle physics and for nuclear physics. The next generation big particle accelerator for nuclear physics is called the Electron-Ion Collider. It's at Brookhaven National Lab in New York. It was approved about one or two years ago by the Department of Energy and is the forefront project for the future of nuclear physics. Studying electron-ion collisions helps us, for example, understand something we can't observe, what the physics was like in the early universe, what the physics is like at very, very short distances, and so forth. We've never collided electrons and ions before, and this big facility is being built at Brookhaven. It'll require a detector that has similarities to detectors that other facilities, but also is rather specialized because it has electrons colliding on ions. The design of that detector is the problem that I'm going to address, and I show it on this slide. In using AI, we don't have a detector design. This is ongoing, and I'm just describing what's being done. In trying to see how to build a detector for it, AI selects—I'm simplifying—a proposed detector configuration. That configuration is then simulated using a big code that's been built over generations, which has everything we know about the physics of nuclear physics and particle physics in its software. That is used to simulate particle collisions in this environment. The reconstruction software measures the detector performance throughout the design stage then. AI comes in by constructing a surrogate model that predicts the performance through the design space. It selects the next configuration where either better performance or important uncertainties are reduced. The cycle is repeated until AI helps us identify a near optimal design. That's more or less the program that's going on. I'm not going to show you any more of that. I'm going to talk about where that is compared to what we mean by a world model. Although this effort is certainly in the spirit of a world model, it uses AI to navigate a human-defined design space. It uses GEANT4, which is a fancy complete simulation program used for particle and nuclear physics that supplies the model of the physical world. It's using both simulations and a human-defined design space. A true world model, AI would internalize the detector's geometry, particle interactions, its electronics, its reconstruction, operating conditions, engineering like cost and engineering constraints, and then simulate the consequences internally, reason causally, identify where its predictions are uncertain, and potentially formulate a genuinely new and fantastic detector concept. That's what a true world model would do. We're not nearly there yet. The present work is an important step toward that. It's a model-based scientific design, but it's not actually an AI-guided optimization. It's actually an AI-guided optimization that's strongly coupled to a physics simulator. It's not yet what we might call an autonomous world model system. That's where we are. You'll hear more about that as it develops. I think it's probably the biggest experimental lab for developing a world model in a complex technical situation. I want to go on to my second topic, which is the topic of discovery in physics and astronomy. The same applies. I just show here that we strongly use probability when we do that. I'm going to try to walk you through that a little bit and show you what the challenge is for AI. Discovery, we should start with the beginning of doing science in a way where we could make a discovery and believe it, and how we establish that. That started in the 15th and 16th century. Actually, modern science started in the 15th or 16th century. Copernicus, at that point, showed us the importance of using mathematics to describe and determine physical phenomena, including the earth going around the sun, if you recall. Galileo showed us that nature itself decides scientific questions through observation and experiment. He, for example, looked through a telescope in 1609 for the first time and saw four moons of Jupiter. Kepler extracted the mathematical laws, and he used it, for example, to show us that the orbit of planets around the sun are elliptical, and then all the details of what that means. Newton, in addition to giving us gravitation, which is the beginning of my research subject, gave us the scientific method itself. He wrote down the scientific method, which we've all kind of seen somewhere, and I want to reflect on that a little bit, because this was the beginning of when we had a method to try to confidently make conclusions about scientific conjectures or measurements or observations. The method in the early days kind of looked like I show here. We observed a phenomenon. We formulated from that, or theoretically, a hypothesis. We derived quantitative predictions, designed experiments and observations to test those predictions, analyzed the data, revised or replaced the theory as necessary, and then tried to confirm or repeat the results independently to gain confidence that they were right. More or less, this was the beginning of actually having some way that made us believe scientific results by following this type of procedure for the next couple of centuries. It was really the last century where things changed. In the last century, we developed statistics as a subject and statistics as a method of measuring or determining whether what we determine scientifically is correct or not. In a modern day version of the scientific method I show here, we observe a phenomenon or develop a theory. We formulate a hypothesis then, derive quantitative predictions, design experiments, and observe to test those predictions. That all follows the original methods. Then we analyze the data, and we establish, typically, a confidence level on the result that we have. That's actually integrating modern statistical techniques into what we do. Over a period of time, we've set as a standard a discovery level of five sigma as where we believe a scientific result. When something new happens and it's published in the New York Times, even not just a scientific journal, they may say it's established five sigma, or it's only three sigma in the newspaper article. If it's only three sigma, we're waiting to see whether it holds up, but this is the result so far. This then has to be confirmed, any scientific result, as we saw. Now we want to replace that somehow by putting in machine learning. The first steps will be the same. Then we have to do something equivalent to the probability or establish a score. The discovery level then must be established in a reliable way, and then revise or replace the theory as necessary. The reason we have to do this is that, basically, machine learning is a black box. We have to determine what it is. We can't use statistics. We have to replace statistics with some way of determining with confidence the results that we get from machine learning or from AI. Let me reflect then on what we're doing. If we have a statistical result and it's purely statistical without other effects, then we've established with experience a five sigma level where we believe the statistics mathematically that equates to a probability value of 310 to the minus seven or a probability of one in 3.5 million chances that the result is random standard statistical fluctuation or a fluke. The ironclad rule used by physicists to declare this then has become our discovery level. I'll show you two examples and two of the most important discoveries of the last decade. First is the discovery of the Higgs boson. The Higgs boson was discovered at CERN on the large particle accelerator. It's named after one of the inventors of the theory that explains and describes mathematically and physically how Einstein's E equal mc squared works, how you make mass from energy or energy from mass. The production of particles of having mass, how they achieve their mass, basically comes from this theory. The theory is something like 70 years old. In it, one of the inventors of the theory recognized that a particle would be involved in addition to just the theory being a theoretical way to explain how mass is generated from energy and that a new particle would be there. The details of where that particle would be were not predicted, but its characteristics were. If we found it, we know what its characteristics are. I won't go through that in detail because that doesn't matter here, but this graph does. This is the discovery graph that was published in the literature in 2012, which shows a little bump over a statistical background, a physics background actually, and how significant is this bump. This is the bump after it reached five sigma significance. It was seen to be observed several years before, and they kept taking data, and the sigma grew. There were two experiments, and on the right plot, you can see that the two each achieved more than five sigma. At that point, they announced the discovery, and that discovery won a Nobel Prize for Higgs and its collaborators. The experiment has gone on to look at the properties of this Higgs boson and whether there's any more Higgs and other characteristics. The second is the discovery that I was part of, the discovery of gravitational waves. The plot on the left is the one in the discovery article. The top row shows the experimental data as a function of time. The time goes from roughly two-tenths of a second from left to right. The plot on the upper left is the detector in the state of Washington, Hanford, Washington, and upper right in Livingston, Louisiana. You can see they look pretty much alike. If you go down one level, the next is the predictions using Einstein's theory of general relativity, which you can see is a smoother version of the same plot. Smoother because the ones on the top have some noise still that shows in the data. If you subtract one from the other, you get the third level, which is just random noise. It was key to look at that to see that all that's there is random noise. So we see Einstein's equations superimposed on random noise, and that's the signal from gravitational waves. That was interpreted then through Einstein's equations as having been made by the collision of two black holes. You can see over on the right, one of them 36 times the mass of our sun, a second one 29 times the mass of our sun, with some errors. The final black hole after it merged, which we also measure, was 62 times the mass of our sun. You'll notice the difference between those is a few solar masses, a few times the mass of our sun, of our earth. Our sun changed to energy, and at that moment in time, that was the biggest, brightest point in the whole sky. It occurred 100, 1.400 megaparsecs, which is a billion light years away. A billion light years away means it happened about a little over one billion light years ago. When this event happened, it was exactly during the period when on earth here, we were just evolving from single cell to multi-cell life. Now, a billion years later, we've evolved to us. It's, in the meantime, traveled at the speed of light and reached us, and we were ready to detect it and observe it. The next slide is the background distribution showing that that again reached the five-signal level, which meant that the false alarm rate, the chance that we were not seeing what we saw from the statistical point of view, was once per 203,000 years, giving us great confidence that we saw what we thought we were seeing. Now, let me address how and what the challenge is for neural networks to do the same thing. We want to use neural networks and have the same story that I've just told you, a Gaussian significance, where we can tell confidently that we make a discovery. I talk about that here. We are, when we use machine learning, we do have ways to measure a confidence level. For example, a neural net, a boosted decision tree or transformer produces significance giving a score. The score can be something like, say, 0.99, and we don't have a false positive rate. Neural networks we know in doing this tend to be overly confident. Most things that we do that aren't right add to seeing false events. Calibration techniques are used to take care of this problem. For example, Plath scaling in order to reduce the overconfidence. Although these techniques tend to improve the score, they still are not reliable, certainly at the level of making a high significant five sigma discovery about the physical world or some other application. The challenge then is to develop the equivalent confidence using ML to that that we have using statistical techniques. Developing a reliable determination for a machine learning result, equivalent to a statistical analysis, is presently the big challenge in order to apply AI to making scientific discovery. I want to point out that this poses a problem. I'll tell you where we are in solving it in just a second, but I point out that this is not just a specialized problem for physics and astronomy. If we want to take results from machine learning and AI with enough confidence to apply them broadly to some important problems, we need confidence. I'll give two examples. To apply AI significantly to making diagnosis or understanding things medically and making applications in medicine, we need to develop a level of confidence that lets us do that, whether we're analyzing x-rays or whatever it is. Secondly, the whole defense industry also needs confidence in what their weapons are going to do or not do in the defense mechanisms. So applications to a huge part of our society needs to do the same kind of developmental work that we need to for discovery physics. The approaches that we have, I can list a few. Null simulation or permutation tests, this gives classical p-value or sigma significance for an ML-derived statistic. It's potentially very reliable if you do all the pieces of it right. Sample splitting and blind analysis, protection against training or adapting to or discovery data, that's also reliable. The reliability is very high, assuming that the rules are rigorously enforced. Simulation-based inference, where we have a posterior distributions or likelihood ratio estimates from complex simulators. That, again, the reliability is powerful, but requires extensive calibration. And there's Bayesian neural nets and so forth that I'm not going to talk about everything. The best current solution I'm kind of outlining here is to treat a neural network as a test static generator. Calibrate its output on independent data or simulated samples under null hypothesis. Determine how often background itself produces an equally extreme result. That frequency then gives a p-value, which can be converted to an equivalent Gaussian significance. The reliability doing this can be high if the network output is treated like any other statistic and its null distribution is calibrated independently. But there's many pitfalls. Using the same data twice is a common thing that you have the data and use it here and use it there. Simulation mismatch is a problem. Extreme tail scarcity, that is, understanding what is in a tail. Overtraining. I'm not going to define each of these. I'm just giving you the labels. Look elsewhere effect, that is, when you look somewhere you don't see something, you look somewhere else, you're more likely to see it than statistics tells you because you've biased yourself. Systematic uncertainties, poor calibrated scores, et cetera, et cetera. So this is the challenge and this is where I'll end in my talk. We want to start from experimental data, put it into an AI model, like I'm showing on the left, and get some version of an AI statistic that we can believe, which requires going through the steps, which I've kind of alluded to in the previous slides, in a set of steps, all of which have a variety of possible failure modes and problems that have to be done reliably. So, of course, this is much more complex than just showing that you have a Gaussian distribution in statistics, but, of course, the value of being able to use AI and machine learning to use it directly in physics, experiments, medicine, defense, and so forth. So this is just an outline of what it is. I'm not going to go through that in detail. AI then ranks the evidence. Independent statistical validation establishes that it's a discovery. And so this is the program, and I think in principle it's all doable, but we're not quite there yet. Thank you.