AI Not Replacing Cosmologists: Why Human Knowledge is Still Needed

The challenge in exploring the cosmos is its vastness, making it difficult to interact with the distant stars. Consequently, we rely on observable data to test our hypotheses on the creation of galaxies.

Computer simulation of these celestial bodies has proven to be an extremely useful tool in wrapping our heads around the nature of reality, and as Andrew Pontzen explains in his new book The Universe in a Box: Simulations and the Quest to Code the Cosmos, recent advances in supercomputer technology continue to revolutionize our ability to model the complexities of the cosmos (not to mention the myriad of Earth’s of challenges) on a smaller scale. In the excerpt below, Pontzen looks at the recent crop of astronomy-focused artificial intelligence systems, what they can accomplish for the industry, and why he’s not too worried about losing his job to one.

Adapted from THE UNIVERSE IN A BOX: Simulations and the Quest to Code the Cosmos by Andrew Pontzen on June 13, 2023, published by Riverhead, an imprint of Penguin Publishing Group, a division of Penguin Random House LLC. Copyright © 2023 Andrew Pontzen.

As a cosmologist, I spend much of my time working with supercomputers and creating simulations of the universe to compare with data from real telescopes. The goal is to understand the effect of mysterious substances such as dark matter, but no human can digest all the information recorded about the universe and all the results of simulations. This is why artificial intelligence and machine learning are a key part of cosmologists’ work.

Consider the Vera Rubin Observatory, a giant telescope built atop a mountain in Chile designed to repeatedly image the sky over the next decade. It doesn’t just build a static image: it specifically looks for objects that move (asteroids and comets) or change brightness (flashing stars, quasars and supernovae) as part of our ongoing campaign to understand the ever-changing cosmos. Machine learning can be trained to detect these objects, allowing them to be studied with other, more specialized telescopes. Similar techniques may even help sift through the changing brightness of vast numbers of stars and find signs of host planets, furthering the search for life in the universe. In addition to astronomy, there is no shortage of scientific applications: for example, Google’s artificial intelligence subsidiary DeepMind has built a network that can surpass all known techniques to predict the shapes of proteins starting from their molecular structure. This is a crucial and difficult step in understanding many biological things. processes.

These examples illustrate why scientific excitement around machine learning has grown over the course of this century, and strong arguments have been made that we are witnessing a scientific revolution. Back in 2008, Chris Anderson wrote an article in Wired magazine declaring the scientific method, in which people propose and test certain hypotheses, to be obsolete: “We can stop looking for patterns. We can analyze data without hypotheses about what it might look like. We can throw numbers into the largest computing clusters , that the world has ever seen, and allows statistical algorithms to find patterns that science cannot.

I think this is taking things too far. Machine learning can simplify and improve certain aspects of traditional scientific approaches, especially when processing complex data is required. Or it can digest text and answer factual questions, as demonstrated by systems like ChatGPT. But it cannot completely displace scientific reasoning because it is all about seeking a better understanding of the universe around us. Finding new patterns in data or repeating existing facts are only narrow aspects of that search. There is a long way to go before machines can do meaningful science without human supervision.

To understand the importance of context and understanding in science, consider the OPERA experiment, which in 2011 apparently determined that neutrinos travel faster than the speed of light. The argument is close to blasphemy in physics, since relativity would have to be rewritten; the speed limit is an integral part of its design. Given the enormous weight of the experimental evidence supporting relativity, questioning its foundations is no light step.

Knowing this, theoretical physicists lined up to reject the result, suspecting that neutrinos must really be traveling slower than reported. However, no problem was found in the measurement – until six months later, OPERA announced that the cable had come loose during their experiment, which explains the discrepancy. Neutrinos did not travel faster than light; information suggesting otherwise was incorrect.

Surprising information can lead to revelation under the right circumstances. The planet Neptune was discovered when astronomers noticed something wrong with the orbits of other planets. But if the claim contradicts existing theories, it is much more likely that the data is flawed; This was the feeling physicists relied on when they saw the OPERA results. Such a reaction is difficult to formulate into a simple programming rule as computer intelligence because it is halfway between the worlds of information retrieval and pattern search.

Machines will not copy the human elements of science unless they can integrate their flexible data processing into a larger data set. There is an explosion of different approaches towards this goal, driven in part by the commercial need for computer intelligence to explain its decisions. If a machine in Europe makes a decision that affects you personally—perhaps rejecting your mortgage application or raising your insurance premiums or pulling you over at the airport—you have a legal right to ask for an explanation. This explanation must necessarily extend beyond the narrow world of data in order to relate to the human perception of what is reasonable or unreasonable.

The problem is that it is often not possible to create a complete explanation of how machine learning systems make a particular decision. They use many different types of data, combining them in complex ways; The only really accurate description is to write down the computer code and show how the machine is trained. It’s accurate, but not very explanatory. At the other extreme, you could point to an obvious factor that controlled the machine’s decision: maybe you’re a lifelong smoker, and other lifelong smokers died young, so you’ve been denied life insurance. This is a more useful explanation, but it may not be very accurate: other smokers with different work histories and medical histories are admitted, so what’s the difference? A fruitful explanation of decisions requires a balance between accuracy and comprehensibility.

In the case of physics, the use of machines to create fluid, precise explanations anchored in existing laws and frameworks is in its infancy. It starts with the same requirements as commercial AI: the machine not only needs to demonstrate its decision (that it has discovered a new supernova, for example), but also provide a small, digestible amount of information about why it has made the decision. This way you can begin to understand what is in the data that has led to a certain conclusion and see if it agrees with your existing thoughts and theories of cause and effect. This approach has begun to bear fruit, yielding simple but useful insights into quantum mechanics, language theory, and (my own collaboration) cosmology.

All these applications are still framed and interpreted by humans. Could we imagine instead that a computer formulates its own scientific hypotheses, balances new information with the weight of existing theories, and explains its findings by writing a scientific paper without human help? This is not Anderson’s vision of a theory-free future for science, but a more exciting, disturbing, and far more difficult goal: machines to build and test new theories on top of hundreds of years of human vision.