How long have astronomers been using artificial intelligence? How do computers sift through massive amounts of data? Will AI discover something new about the universe? I discuss these questions and more in today’s Ask a Spaceman!

This episode is sponsored by BetterHelp. Give online therapy a try at betterhelp.com/spaceman and get on your way to being your best self. Visit BetterHelp to get 10% off your first month!

Support the show: http://www.patreon.com/pmsutter
All episodes: http://www.AskASpaceman.com

Follow on Twitter: http://www.twitter.com/PaulMattSutter
Like on Facebook: http://www.facebook.com/PaulMattSutter
Watch on YouTube: http://www.youtube.com/PaulMSutter

Read a book: http://www.pmsutter/book

Keep those questions about space, science, astronomy, astrophysics, physics, and cosmology coming to #AskASpaceman for COMPLETE KNOWLEDGE OF TIME AND SPACE!

Big thanks to my top Patreon supporters this month: Justin G, Chris L, Barbara K, Duncan M, Corey D, Justin Z, Naila, Scott M, Rob H, Justin, Louis M, John W, Alexis, Gilbert M, Joshua, John S, Thomas D, Simon G, Erin J, Valerie H, David B, Tim R, Tom Van S, Mark R, Alan B, Craig B, Richard K, Dave L, Stephen M, Maureen R, Stace J, Neil P, COTFM, Stephen S, Ken L, Alberto M, Matt C, Joe R, David P, Ulfert B, Sean M, Edward K, Tracy F, Sarah K, Steven S, Ryan L, Ella F, Richard S, Sam R, Thomas K, James C, Jorg D, R Larche, Syamkumar M, John S, Fred S, Homer V, Mark D, Colin B, Bruce A, Steven M, Brent B, Bill E, Tim Z, Thomas W, Linda C, David W, Aissa F, Marc H, Avery P, Scott M, Thomas H, Farshad A, Matthias S, Kenneth D, Maureen R, Michael W, Scott W, David W, Neuterdude, Cha0sKami, Robert C, Robert B, Gary K, Stephen J, dhr18, Anna V, Matthew G, Paul & Giulia S, Steven M, Louis M, Michael C, Alyssa K, Lode D, Roger, Bob C, Patti H, Red B, Benjamin M, BlueDragon, Stephen A, Ian S, James R, Skip M, Robert O, Adam I, Lynn D, Jeffrey C, Allen E, Paul G, Michael S, Jordan, Colin H, Jessica M, Thomas H, Reinaldo A, Amy Z, Adam I, Sheryl, Skip M, David W, Denis, Sue T, and Jonathan S!

Thanks to Cathy Rinella for editing.

Hosted by Paul M. Sutter.

 

EPISODE TRANSCRIPT (AUTO-GENERATED)

I'll stay up front that I had a hard time structuring this episode. I went back and forth of just just how to put this together. The topic of today is the use of artificial intelligence in astronomy of which there is, let me see, a lot. But artificial intelligence or AI is this really broad concept that has not appeared in previous Ask A Spaceman episodes. So I can't just say, well, listen to episode blah blah blah and then dive right in.

So so we're going to have a bit of a split personality today. Half the episode is about AI and specifically the kind of AI routinely used in astronomical research, and the other half is about all the research itself. And, of course, I have my own personal thoughts and feelings on the matter, which will be sprinkled here and there as usual. But to get us started, I want you to imagine having a friend. Insert joke about this sounding ridiculous.

I mean, who has time for friends when there's so much universe to study, but imagine having a friend in this not any friend, a certain kind of friend. Maybe you already know somebody like this. Someone very, very smart. Someone who is able to pick up new information quickly, who can learn new languages like like that, who absorbs complex topics on a variety of subjects with ease, just like a mental sponge. And this friend loves to talk.

Typical professor attitude. You ask this friend a question and you're guaranteed to get a thorough, detailed, lengthy answer. But this friend has one fatal flaw. This friend doesn't always get it right and what's worse, this friend doesn't know when they're getting it wrong. I'm reminded of a phrase to describe some people and almost certainly never used to describe me and the phrase is often wrong, never in doubt.

For example, your friend who is very intelligent and can pick up languages and learn things quickly will learn French like that in a weekend. Yeah. Yeah. I learned French. There you go.

Boom. Done. And if you ask them to spout off a French phrase, they will and they'll be confident about it. They will look you right in the eye and they will rattle something off like it's second nature. But because you don't speak French, you don't know if they're saying something witty, intelligent.

You don't know if they're quoting a passage from famous French literature or if they're saying This friend is often wrong, but never in doubt. It turns out we all now have a friend like this and that friend is artificial intelligence. Now we need to take a step back to set some things up. Artificial intelligence AI is an incredibly broad term that encompasses a whole host of programs, research, ideas, software, hardware, or algorithms, you name it. Some aspects of AI research have become so mainstream that they don't even qualify as AI anymore.

Like like image processing. Seriously, developing the ability for a computer to capture light, process it, store it as digital information, display it again was originally considered AI research because that's something that intelligent creatures do is they absorb infrared radiation hitting sensors and then process and store it later. So that was considered AI research, and now it's just something that our phones do. So many aspects of AI research is just a part of everyday life when we don't even think about it. And on the other end of the spectrum is the holy grail of AI research, which is to develop a wholly sentient or self aware computer or at least a machine that can mimic human intelligence to such a degree that it's indistinguishable from 1, and yes, there is a significant debate about whether that difference matters.

This kind of research falls under the general label of, well, general artificial intelligence. Yep. If you think of the evil version of general AI, those those are the terminators, and they will not be the subject of today's episode. And then there's a whole bunch of cool stuff in between. The interesting lines of research that touch on aspects of intelligence without attempting to or even claiming to attempt to be generally intelligent.

We're not trying to make a thinking computer. We're just trying to mimic some aspect of human cognition or intelligence. As a bit of personal background, before I began studying physics, I was actually a computer science major. And I got to see, although not participate in, because I was just just yay big. Some of this research myself and through the years I've kept my eye on the progress in the field and especially how that progress intersects with physics and astronomy.

Where once I became a physicist, I quickly learned that physicists and astronomers are just amateur computer programmers anyway, but that's another episode. The AI research and its applications to science or medicine or everyday life has been growing in academic circles for some decades. One of the most popular fields of AI research, which remember AI is this broad umbrella that encompasses everything related to to making computers kinda smart and mimic human intelligence. One of the most popular fields is a general strategy called machine learning, where we get computers to learn new things and then use that knowledge to perform new tasks. So traditional computer programming is you program the computer, you tell it what to do, and then it does the thing.

With machine learning, instead, you teach the computer how to learn. That's what you program in is an ability for the computer to gather new knowledge, then you give it new knowledge, and then now it has acquired new knowledge, it can go perform some task. Hopefully, some task that is more sophisticated or more complicated or outright impossible to do with traditional computer programming where you just tell the computer what to do in the first place. Under this umbrella of machine learning where we get computers to learn new things and then have them perform tasks, there are a set or there is a set of computer algorithms called neural networks where we have computer software that mimics how the brain learns. And here's a hint, it uses a bunch of neurons.

So this is where we are with, and why I'm highlighting this because it's specifically neural networks that intersect astronomical research a lot in neural networks are a kind of machine learning, which is a branch of artificial intelligence research. The show is not titled Ask a Computer Man, so I will most definitely not get into the nuts and bolts of how these algorithms work. But I will say that you should keep in mind that these are not actual neurons when someone talks about neural networks or machine learning. This is not an actual brain reproduced in silicon or in microchips. What we're looking at is biological analogies like a brain.

A brain has a bunch of nodes called neurons. These neurons connect to each other. And when the brain learns something, the connections between certain sets of neurons strengthens. And with those connections strengthened, that knowledge is acquired by the brain and then you go out to do something. If I throw a ball at you a bunch of times and you and you practice catching, catching, catching, there are certain sets of neurons in your brain that get their connections reinforced.

And then the next time I throw a ball at you, boom, you can do it without thinking. The act of me speaking to you right now, I have certain connections in my brain with that encodes this knowledge in a way we don't understand, but it encodes this knowledge. I'm speaking to you and you are learning it. Hopefully, those same connections are getting reinforced in your brain, your your learning. So the point of neural networks is to simulate this in software.

We don't have actual neurons connecting to each other, but we have software based imaginary neurons that build imaginary connections, and then this software can learn. I know it sounds kind of freaky and it kind of is, it's also kind of cool. The end result is that we can teach these algorithms new things and then we turn those algorithms around to perform some interesting task in the real world. We are teaching computers. This kind of AI research has exploded in popularity in recent months with the release of chat GPT, which is an AI algorithm that can produce reasonable sounding and moderately coherent text based on prompts.

ChatGPT represents a kind of AI machine learning neural network known as generative AI, that's the g, where a computer algorithm creates output that mimics some expression of human intelligence, in this case, text. Algorithms like chat GPT aren't actually intelligent themselves the same way we're intelligent. We actually understand the words that we are learning and the meaning chat GPT just emulates that. It mimics that in a very, very faithful way. They aren't intelligent, but they sure do sound like it.

And this is our friend. This is our friend who we are going to enlist to do astronomical research. Our very intelligent friend who is sometimes wrong, but never in doubt and doesn't know when they're wrong, but is always very, very confident. Our friend is a neural network. Now, to understand how our friend, the neural network, works and how it might be of use to your average astronomer, let me give you an example.

Let's say you want to use one of these sophisticated neural network machine learning AI algorithms to get a computer to identify cats. You can imagine this is a very difficult thing to just tell a computer what to do. It's hard to encode what a cat looks like in any given picture. If I just show you a picture of a cat and I say, okay, like how do you instruct a computer to identify a cat? Well, one way is to do it through neural networks.

You feed it a bunch of pictures. You train this neural network. You train the computer algorithm. This is the learning phase with a bunch of data. You give it tons of pictures of cats.

The Internet is full of them. This is not a difficult task. Different colors of cats, different sizes, different poses, different angles, different times of day, different backgrounds. Give it all sorts of cats. By the way, if you've ever had to fill out one of those CAPTCHAs, you know, you enter your password and says prove you're a human and you have to pick out all the all the motorcycles, all the street lights.

What you're doing is providing training data for an artificial intelligence system. Anyway You feed this information into the neural network into this algorithm the neural network learns and I'm using scare quotes here What a cat is by pulling commonalities and connections out from all those pictures. For example, some of these commonalities and connections may be obvious. Cats tend to have 2 pointy ears. Cats tend to have long tails.

Cats tend to have vertically oriented eyes. They tend to have 4 legs and varying amounts of fur. There are some things we can point to. If I gave a human a hundred pictures of cats and they're all random cats at different times a day, then you a human could describe what's common between them. Like, oh, yeah.

Yeah. I see mostly pointy ears, mostly four legs, mostly tails that tend to curl around. There are some things we could say out loud and understand with each other to define a cat. But these algorithms, and this is crucial when it comes to trusting their output, you have a very intelligent friend that can very quickly pick up what is a cat based on a 1,000 pictures of cats. This friend develops an intuitive sense of what a cat is.

But some of these connections and commonalities aren't necessarily obvious and don't necessarily make sense. These are connections, an essence of catness that I can't even describe because sometimes they're impossible to describe. You know, this gets into this actually, like, deep philosophical question here, like, what is a cat? And how do you know it's a cat and when it's not a cat? How do you know?

How can I draw a stick figure of a cat and show it to you without telling you what it is and you'll be like, yeah, I think you drew a cat? How do you know? Is it the pointy ears? Well, not all cats have pointy ears. Is it the long tail?

Not all cats have long tails. Is it the 4 legs? Well, if a cat is missing a leg, you'd all of a sudden does don't say if you see a cat with 3 legs, you know, poor kitty that's been amputated, you don't sit there and say, oh my god. What is this strange three legged creature? No.

You say, okay. That's a cat missing a leg. One of the wildest examples is if you see a cat behind a fence and you only see parts of the cat, you don't freak out saying, what is this multi structured creature that's segmented in different parts? No. Immediately, you say that is a cat standing behind a fence.

There are some things about catness that are obvious and some things that are not, some things we can't put into words. And the spooky and cool part of the about these neural network algorithms is that they capture that catness. They look at thousands of pictures of cats, and they build up some sense of catness that the algorithms themselves can't fully describe. We'll get back to that in detail later. That's a very important nugget.

Instead, the essence of katniss is stored in the connections between these artificial simulated neurons. For the nerds among you, that's something called latent space. And the neural network, through these neural connections, is able to establish catness that it can't put into words, I can't put into words, and that's kind of how the brain works and how we learn new information. We can identify cats without precisely saying entirely how we do it. So mission accomplished in mimicking human behavior.

So our good friend, our very smart friend, this friend has never encountered cats before. And we show our friend a 1,000 pictures, and we say, every one of these pictures has a cat in it. And the friend scans through all the pictures and say, I understand cats. The algorithm is trained. The algorithm has found all these interesting and non obvious connections and commonalities that give it a general picture of what a cat is, And now we can point this algorithm at the real world and give it random pictures and say, okay.

Do you see a cat in here or not? And it uses its knowledge of what makes a cat a cat, and it gives you an answer. This picture has a cat. This picture does not have a cat. This picture has a cat.

No. No. No. Yes. Yes.

Yes. And because it's a computer, it can do this very very quickly. You can process thousands of pictures and know with the speed that only computers are capable of which ones have cats in them and which ones don't. We can find all the pictures in the world and categorize them as either having a cat in them or not having a cat in them without once needing a human to look at it. 8, you can expand this by labeling the cats that you train your friend on, you train this neural network algorithm on, instead of just saying, oh, there's cats in every one of these, identify the essential features of catness and use it to identify cats in the real world, you can say, okay.

Okay. Here are a bunch of pictures of tuxedo cats. Here are a bunch of pictures of tabby cats, Siamese cats, sphinx cats. Oh, so weird. On and on and on.

And now you have an automated cat classifier because it doesn't the algorithm doesn't just pull out catness. It says, okay. I understand tuxedo cats. I understand tabby cats. I understand Siamese cats.

And I know the differences between them. And the algorithm, your friends can say, give me any picture you want. I will tell you what kind of cat it is. Neural networks, in other words, are great for deciding categories. I may or may not have structured the entire episode up to this point in order to make that joke.

I will not blame you for not contributing to Patreon this month. If you're wondering how chat GPT works I don't know how I'm gonna recover from that. If you're wondering how chat GPT work works, it learned how to write text because we shoved pretty much the entire Internet down its digital throat. And from there, it found the connections between words, what words commonly follow other words, and so it can spit it out. And it took a lot of massaging.

This is not an easy process, and it can put together strings of words that sound good together. It's a very very smart friend that was trained on text, not pictures of cats. Of course, we have more interesting problems than whether pictures have cats in them or not or what kinds of cats are in the pictures. The bottom line is this, machine learning algorithms and neural networks in particular are very, very good at finding deep and hidden connections in datasets, which makes them very, very fast at processing large quantities of new data and discovering new information, which is the foundation of a data heavy science like astronomy. Astronomy has always always run on massive amounts of data and always run on doing one of 3 things either looking for patterns or looking for anomalies or categorizing objects or events with the hopes of further understanding.

For example, Kepler. Kepler discovers that the planets move in ellipses. How was he able to discover this? With massive, massive amounts of data. Books filled with handwritten tables of the positions of the planets going back years, decades.

Massive amounts of data. A handwritten spreadsheet that he manually computed possible orbital fits to figure out what shape these orbits were until he found ellipses. Previous astronomers didn't figure it out because they didn't have the data. Or look at William Herschel who discovered Uranus. How did he find it?

By collecting massive amounts of data and looking for anomalies. He was just doing normal sky surveys, and he was able to do enough of them repeatedly over the same patch of sky night after night after night where he was able to find one of these stars moving and stars aren't supposed to move, he discovered a new planet. Previous astronomers had spotted Uranus. Galileo may have spotted Uranus and recorded it, but they didn't have the data to realize that it was an anomaly, that it was a planet. Or our good friend, Annie Jump Cannon, who is able to finally come up with a sane categorization scheme for stars based on their spectrum, based on their colors.

How was she able to do it? By having a 100,000 spectra at her disposal. Previous astronomers couldn't do it because they didn't have enough data. And once she had a sane categorization scheme, we were able to develop the theory of how stars evolve. Artificial intelligence, specifically machine learning, specifically specifically neural networks, can help with all of this.

Once you have a very smart friend who is able to learn things very, very quickly and able to discover deep and hidden connections in the data, they they can do anything you want. They're very smart, and they're very quick about it. You wanna look for anomalies? Okay. You train your algorithm on cats, and then you say, show me all the weird cats in the real data.

Now that you have an understanding of catness, go out in the real world. I will show you a bunch of pictures of cats. And if anything kind of sort of looks like a cat, but isn't quite, doesn't quite fit the bill, alert me. And then you go to bed. You eat a snack.

You do something else, and you come back. And your good friend, the neural network, has produced a list of things that kinda sorta look like cats but aren't quite maybe cats. Maybe it discovers dogs. Maybe it discovers lynxes or lions. It finds anomalies in the data.

Or categorization. You tell your good friend, the neural network, here are the different kinds of cats. Now now that you're trained, now that you're very smart, go out in the real world, and here's not 10,000 pictures. Here's a 100,000,000 pictures of cats. I need you to categorize all of these.

Boom. Boom. Boom. I need you to do it very quickly. Any jump cannon categorized her spectra by hand.

What she did over the course of decades would take a computer a few minutes, would take a neural network a few minutes. That's fast. You wanna look for patterns? Neural network, your good friend, is right there ready for you. Neural networks can accelerate astronomy because they go deeper into datasets than humans do, and they can do it faster, which is a win win.

Here are some examples, and this is just perusing some of the latest papers. Seriously, these are papers written in the past 2 weeks from the recording of this episode. There's a paper on using AI for galaxy classification. You know, classifying galaxies is mostly done by hand. When we say, oh, that's a spiral galaxy.

That's a elliptical galaxy. That's a lenticular galaxy. And then there's subclasses. That's a type 2 c on the Hubble fork diagram blah blah blah. That's all done by hand.

Neural networks can do it like that. Blazar classification. Not just objects, but events. We we see explosions in the sky. Sometimes blazars are are certain wavelength range, sometimes certain frequencies, sometimes certain durations.

Boom boom boom. Let's classify them. Let's get this over. Let's have it all automated. So we don't need a human involved in the loop.

You know, some poor grad student stuck in some underground bunker of a laboratory. Been there. Manually classifying these events or desperately trying to come up with a traditional computer algorithm and trying to capture all the nuance of these different classifications, things that humans just intuitively get and understand, have a neural network do it instead. Have our good friend do it instead. Asteroid identification.

Hey. Scan through all these gigabytes, terabytes, petabytes of astronomical data. Tell me which ones are the asteroids. Boom. A trained algorithm can sift through enormous amounts of data and do all this work far faster than a human can.

We have a very, very good friend who's very smart. Perhaps the most powerful application of machine learning is in cosmology, and that's because in cosmology, we can't actually see the parts of the universe that we're interested in looking at. When we perform our studies and our observations, we're limited to what gives off light, galaxies, the positions of galaxies, the shapes of galaxies, the types of galaxies, the arrangement of galaxies. If we're lucky, we have something like the cosmic glycoide background, but that's still made of light. And the stuff that gives off light, that interacts with light, is less than 5% of the energy budget of the entire universe.

Most of it is hidden in the form of dark matter and dark energy. Dark matter and dark energy are the most important parts of the universe and we can't directly observe them. There is so much that is hidden from us. The amount of dark matter in the universe that is hidden from us from direct observation. The amount of dark energy that is hidden from us from direct observation.

The relationship between dark matter, dark energy, and normal matter. How does different amounts of dark matter and dark energy affect the evolution of galaxies? We can't directly access that. How does the fundamental nature of dark matter and dark energy affect the evolution in history of the cosmos? We can't directly access that.

So our job as cosmologists is to connect the hidden to the scene, to take our slim slice of data and the observations that we have and use that to understand the nature of the mysterious hidden components of the universe. We are trying to grasp at fundamental physics the nature of dark matter, the nature of dark energy, how much dark matter is there in the universe, how much dark energy is there in the universe, has dark energy changed, Is dark matter made of 1 particle, different kinds of particles? Do they interact through new forces of nature? What fundamental laws govern the the earliest moments of the universe? Those are the questions we are trying to answer, but we can only see what we can see.

And so in traditional cosmology, the way we bridge this gap is to use our knowledge of the laws of physics to guess at the fundamental relationships between the hidden and the seen. So for example, we say, okay. The expansion of the universe is governed by Einstein's general theory of relativity. Within that framework, we have ideas like inflation. Structure grows over time.

At first, the dark matter does this. And then the normal matter comes in, and then, it collapses under its own gravitational weight. And then nuclear fusion happens, and then it emits light, and then radiation does this to its surrounding environment. And it leads up to the evolution of a galaxy, and we see a galaxy today. We guess.

We fundamentally guess about what the universe is made of and how it works. We guess about the amount of dark matter and dark energy. We guess about their properties. We guess about the laws of physics. We we also test the laws of physics.

But, fundamentally, we're guessing at it, and then we see how close we get to the real universe. And we use our knowledge of physics, and we express this through computer simulations. I should do a whole episode about how computer simulations accelerate scientific discovery. That'd be a fun episode. Feel free to ask.

We encode. We make mock universes on a computer. We encode the laws of physics. We encode the fundamental ingredients. We cook up a little universe, and then we see how close we get.

One major problem with this is that there are not so enough supporters on Patreon, especially after the cat joke. If you're willing to forgive me, please go to patreon.com/pmsutter. I truly, truly appreciate all the support. I really do. I might make more jokes like that.

I don't know what will happen if I get more Patreon supporters. We'll see. No. The real problem is that our usual tools in physics can only take us so far. One of the reasons that our usual tools when we just take, okay, guess about the ingredients of the universe, take our knowledge of the laws in physics, simulate a universe compared to the real thing.

That gets us pretty far. Don't get me wrong, but it can only get us so far. One of the reasons it can only get us so far is that cosmology is one thing astrophysics is another. Cosmology concerns itself with the big picture. You know, the evolution of the entire universe, the the foundational properties of dark matter, Very, very basic stuff.

But to actually grow a galaxy, you need messy astrophysics. You need magnetic fields, cosmic rays, neutrinos somehow get involved, star formation, supernova going off. All of these messy astrophysical processes, these small scale processes, and, yes, in cosmology, a a star going off is small, change the nature of galaxies and how they evolve. So when we take a picture of a modern day galaxy, like, we take a picture of the Andromeda galaxy, some of the present day properties of Andromeda are due to fundamental cosmology to the amount of dark matter in the universe. But some of it is due to other things like how efficiently stars form or how efficiently galaxies can cool themselves off after supernova detonate or how big of a black hole they have in their center.

All these complicated questions that have no easy answers. And, yeah, it's all rooted in physics that we understand, but the physics, the interactions become so numerous and so complex. It's hard to keep track. And it's hard to separate out what is the fundamental cosmology and what is the interesting, but less fundamental astrophysics. And when all we have, since we don't have pictures of dark matter or dark energy or pictures of the universe when it was one second old, all we have are the pictures of the mature galaxy that contain information about fundamental cosmology, which is what we're after, and messy astrophysics, which some of my colleagues are after, but I'm I'm not particularly after.

And it gets worse. Not only does astrophysics and things like star formation and cosmic rays mix up and obfuscate and hide the information that we're really after, which is the nature of dark matter, dark energy, origins of the universe, all that good stuff. When we zoom out beyond the scale of individual galaxies and we try to, like, look at the cosmic web and the arrangement of galaxies, oh, you we have some, like, statistical measures of the cosmic web. So for example, we can look at, say, the average distance between galaxies. On average, how far apart are galaxies?

The the fundamental constants of the universe, the amount of dark matter, the properties of dark matter, the evolution of dark energy feed this this measure, the average distance between galaxies. Like, you can imagine if you crank up dark energy and you accelerate the expansion of the universe earlier, then on average by the modern day period, galaxies will be farther apart. So this is a measure. This is a way to access that that fundamental cosmological information. The thing is there's not a lot of information to be had in these very simple measures that we've been able to develop.

Even if we've mapped every single galaxy in the entire universe, we wouldn't have enough information from this measurement alone to get better understanding of dark matter and dark energy than we currently have. We have to go deeper. So we have two challenges in cosmology, and I'm highlighting this because neural networks are so are becoming so enormously popular in cosmology. Probably the most popular application of neural networks in machine learning in astronomy is in cosmology, and it's for this reason because it's hard. And this seems like a juicy problem.

Because one, we need to go deeper into the data. We need to find more connections, more commonalities to tease out the nature of dark matter and dark energy, and what's happening in a cosmological scale is being mixed up and hidden by messy astrophysics, and we're we need to disentangle those 2. And neural networks and machine learning can do both. And so over the past couple decades, there's been this growing interest in the use of machine learning and neural networks to get around this. This is the bright idea.

Train a computer, our good friend, the neural network, by feeding it massive amounts of data and let it find, discover all the secret, complicated, non obvious connections between the hidden and the scene. Let it discover. Like, oh, if I change the amount of dark matter by a little bit, it affects some property of galaxies in this particular way. Maybe they're brighter. Maybe they're larger.

Maybe they're more spread out. Maybe star formation is suppressed. Maybe not. Change the amount of dark energy. See what happens.

And we tell the algorithm what is happening. We say, okay, here's this simulated universe with these properties of dark matter, dark energy, star formation, cooling, cosmic rays, neutrinos. Here's another simulation. Here's another. Here's another.

Here's another. Now figure it all out, and then let it loose on the real universe of which there is only one. So we feed the neural network thousands of different possible universes, let it understand the relationship between all these, you know, astrophysics and cosmology and statistical measures and everything, and then let it loose on the real data in the real universe, and it will tell us what universe we live in. The hope is we teach our very, very good friend all the different possible kinds of universes that we could have lived in, but we don't, or we might, and then we give it the real data and it says, okay, yeah, I've looked at all these potential universes and this one is yours with this much dark matter, this much dark energy, by the way, there's 2 different kinds of dark energy. You should know that.

That's kind of important. I just learned that. I learned it on the real data because you trained me too, and I'm very smart. We're going to show our good friend different kinds of cats, and we're gonna label those different kinds of cats, and then we're gonna show it a picture of a cat that we don't know how to classify, that we don't know how to identify and say, what do you think? What kind of cat is this?

It's just a more complicated version of categorizing. We trained neural networks to identify the fundamental nature of cat ness. Now we're training neural networks to identify the fundamental nature of universe ness, so it can tell us it can categorize our universe. That's the promise. The reality is somewhat more complicated, and it's complicated by 2 very serious concerns.

1 is the training data. We're trying to build connections between the hidden and the seen, but by definition, we can't see the hidden parts. So our training data doesn't come from real observations of the real universe because we don't have real observations of dark matter. We can't see it. It's kinda the problem.

Instead, like I mentioned, we rely on simulations. In order for machine learning to work here, we need our simulations of the universe to be as real as possible. Otherwise, we're just training on junk. For example, if I gave our good friend the neural network pictures of only black cats or only tuxedo cats, it doesn't ever learn to recognize an orange tabby cat or or, god forbid, one of those hairless sphinx cats. I still still don't understand those.

They they disturb me at a deep and fundamental level. Our algorithm, our neural network won't know what to do with the tabby cat if it's only been trained on black cats because it will think that black fur is part of the fundamental nature of catnests and won't know what to do with the tabby. They'll get it wrong. If we want our neural network our good friend to be good at categorizing cats We need to give it as many different cats as possible, and we need these to be accurately labeled. When we tell it, hey.

Here's a picture of a black cat. It needs to actually be a black cat. When we showed a picture of a tabby cat, it needs to actually be a tabby cat. We need to get our training data right. Otherwise, we're introducing a bias, a bias that will mess up the results, a bias that will make it difficult to trust our friend.

So in cosmology, there's so much unknown about how galaxies evolve and how dark matter, dark energy, the evolution of the universe, the fundamental stuff we're after, how it interacts and intersects with the actual evolution of galaxies. There's so much we don't know. How efficient is star formation? What is the role of heating from active galactic nuclei, etcetera, etcetera? How efficiently do cosmic rays transport heat?

How does that affect the shape of galaxies? What about galaxy mergers? Oh, that happens a lot. So to work around this, we feed the machine learning algorithms everything. We craft thousands of simulations where we change not only fundamental cosmological properties like the amount of dark matter, but also galaxy evolution properties as well.

Okay. In this simulation, dark matter is the same, but stars form more efficiently because we don't really fundamentally know how efficiently stars form. And then we just kinda cross our fingers and hope that the neural network, our good friend, just figures it all out and learns what is important and what is not important. That's one thing that makes it complicated because if these simulations aren't right, then we're just training it on junk. Our good friend is not being given good data, and if our good friend is not being given good data, we can't trust the outputs.

If we gave it a bunch of pictures of cats, but we mislabeled everything, then when it goes out to identify cats in the real world, it's gonna be junk. The second concern is that we don't know how the algorithm learns. Well, let me say that in a different way. We know how the algorithm proceeds. It is after all a computer program that we wrote, but we don't know what connections the algorithm thinks are important, what weights it gives to certain properties of the data.

And that's like us. That's like humans. We can describe a cat. We can pick out cat like from not cat like objects, but we can't fully explain why we believe something is a cat or not. And sometimes we change our mind and sometimes we're wrong.

Oh, that's a cat. I I didn't realize it. The same is true for neural networks. They can identify cats, and they can be trained well enough that they can identify cats with a high degree of accuracy, but we can't open up the hood to learn why the algorithm believes that this picture is of a tabby cat or a tuxedo cat and this other one isn't. This makes science with machine learning fundamentally difficult because science relies on reproducibility.

And difficult because science relies on reproducibility in tracing our steps, our methodology, and having errors and uncertainty. So imagine the scenario where we create thousands of mock universes. We feed it into the algorithm, the neural network, and then we point it at the real universe. And it after ingesting the real universe, our good friend, the neural network says, uh-huh. This is the kind of universe you live in.

I have been trained well and I am very smart. And now I've seen the real universe and I can tell you with an extreme amount of confidence that this is it. And you can ask, well, how do you know? And the good friend says, I don't I can't tell you that. That makes science hard, because we can't reproduce that.

We can't trace it. We can't have an independent team verify it because the algorithm just says, yeah. Yeah. Yeah. Yeah.

This is how much dark matter there is. This is how old the universe is. This is how important neutrinos are. This is how efficient star formation is. I've categorized your universe based on all my training data.

Thank you very much. And, I'm done. And you say, how certain are you? Well, very certain. Well, in science, we like to have error bars and uncertainties, and we understand that there are limitations of the data and neural networks don't support that, at least not yet.

In other words, our machine learning algorithms will always be confident about their answer, but we're not quite sure when they're right. There's a major debate happening in the astronomical community right now. More and more astronomers are turning to machine learning algorithms, neural networks to do stellar classification, galaxy classification, asteroid identification, fundamental cosmology, but we're having a difficult time folding this into the practice of science as we understand it because we can't just trust the algorithms. You can look up examples from chat GPT all the time. You can say, Chat GPT, give me, you know, a few paragraphs on the nature of, I don't know, cat habits.

And it will give you an answer, but you don't know if the answer is right. In fact, there's a conference that I'm I'll be attending shortly with the literal title of debating the potential of machine learning in astronomical surveys. My personal take on all this, I've seen it grow over the past couple decades. I've never personally dug into machine learning or used it in my own research. I just feel like we haven't been clever enough that when we look at the complexity of the universe, we say, shoot, Can't figure that out.

Let's have a computer do it instead. I feel like there are smarter ways to go about it, and so I've been investigating smarter ways to go about it. But nobody asked my opinion, so I'll just stay over here in my corner. On the other hand, I have friends and colleagues, great friends, great colleagues, and real ones, not the neural network kind, that are very interested in machine learning, neural networks, AI using this in astronomy, trying to find productive uses of trying to answer some of these very fundamental issues of how do we trust the output, how do we reproduce the output, how do we understand the uncertainty of the outputs, how do we prevent bias in all the training data that they're trying to work on it, and good for them. And they're free to express their views on their own podcast.

My takeaway is like everything else associated with AI, it's unclear if these techniques will bring a revolution in how we understand the universe or just fizzle out in a tangled mess of poorly understood algorithms and the realities of a complex universe that we don't understand even with the help of software algorithms to design to mimic how we learn. To be clear, I hope they're useful. It would be great if we can just point a machine, a computer, an algorithm at observational data, and it can tell us what kind of universe we live in. That will accelerate scientific discovery, that will answer some fundamental questions, that will allow us to develop the fear fundamental theories to to break new ground in physics, that would be great. But we're not quite there yet, and it's going to take a lot of research in AI itself how ironic is that for us to fairly judge the utility of AI in researching the universe.

We have our homework. By the way, I asked chatgpt if AI will work in astronomy. After a list of examples of its use, it concluded with AI's ability to handle vast datasets, find patterns, and make predictions has made it a valuable tool for astronomers. It helps streamline many tasks, making the research process more efficient. As technology and AI capabilities continue to advance, the role of AI in astronomy is expected to expand further.

However, it's important to emphasize that AI complements the work of human astronomers rather than replacing them. While AI can handle data processing and routine tasks, human astronomers are essential for formulating hypotheses, designing experiments, interpreting results, making critical decisions, and providing scientific context. The collaboration between AI and human astronomers is likely to lead to exciting discoveries and advances in our understanding of the universe. Like I said, never in doubt. Thank you to Brett t on email for the question that led to today's episode, and thank you to my top Patreon contributors this month.

We got Justin g, Chris l, Barbike, Duncan m, Corey d, Justin z, Nyla, Scott m, Rob h, Justin Lewis m, John w, Alexis Gilbert m, Joshua, John s, Thomas d, Simon g, Aaron j, and Valerie h. That's patreon.com/pmsutter. It is all of your contributions that make this show possible. I can't believe it. Every single month, it is a pleasure to write these episodes.

It is a joy to read them and record them, and I hope episode by episode, we come closer to complete knowledge of time and space. Even if we need a computer to help us out.

Comment