Hello and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. Is there intelligent life in the universe beyond planet Earth?
My guest is working to answer that very question.
Jerry Jang is a graduate researcher at the Berkeley SETI Research Center at the University of California, Berkeley.
Thank you. to aid in processing data from radio telescopes.
Jerry is also doing AI at the Breakthrough Listen Initiative, the most comprehensive and sensitive search for extraterrestrial intelligence in history. session here at GCC19 talking about Breakthrough Listen and their use of artificial intelligence to make sense of the mounds of data they're collecting.
Jerry, thanks for joining the AI Podcast.
My pleasure. Thanks for having me, Noah.
So as we record this, I think your session was earlier today?
That's correct. It was this morning. How'd it go?
It was great. It was a lot of interest. Always glad to see the room full of people.
Yes, it's a good sign. I mean, people are interested in, you know, we were talking... off air before we hit the button that it's not only the work that you're particularly doing and, and you know, others in your field I've gotten to talk to are doing.
But then it's also the method and looking at, you know, just sort of the the sheer science of the method, but then thinking, oh, this could be applied to other situations as well.
And it's... It's fascinating times in the worlds of AI and high-performance computing and all that.
But let's talk about Berkeley SETI and also about Breakthrough Listen.
What is the work that these organizations are doing and what's your role in them absolutely so uh SETI has a long history at Berkeley so SETI at home as you mentioned uh It's based in Berkeley as well.
I'm not actually involved with SETI at home, so I can't really comment on the current operational details.
The name SETI, a lot of people when they hear SETI, they think of the SETI Institute.
The SETI Institute is an institute based in Mountain View.
And they are historically the center of SETI research.
However, SETI, it's actually the name of the field. the search for extraterrestrial intelligence.
Oh, got it. Okay. I'm too stuck in brand names in my head, you know?
Yeah. Brickful Listen and SETI Institute are...
Right. Okay. And I am involved primarily with Breakthrough List.
With Breakthrough List. which is based in Berkeley with the Berkeley SETI Research Center.
And so you're doing AI at the Breakthrough?
That's correct, yes. Okay, and tell us about that.
Yeah, so the data we're working with are primarily in radio frequency.
So a radio telescope collects complex voltages, what we call the raw data. or what the wireless communication industry call IQ data.
So these are complex voltages, complex time series.
And in real time, we take Fourier transforms of that as a function of time.
So we make... what we call spectrograms.
So these are two-dimensional data types.
Essentially we have first the time series, then we have the two-dimensional image. type of data.
So I learned AI primarily from the computer vision community because a lot of the modern techniques are developed there.
And our data is also one or two dimensional.
But there are some differences. My work is to apply a lot of these techniques into our domain with modifications, considering the differences.
And so kind of for the uninitiated, you said you're using a radio telescope.
And so primarily... You're looking for radio transmissions to maybe use a not, Perfectly accurate word, but as out in the universe as signs of something coming from other beings.
Right. Right, right, absolutely. So I guess we should have mentioned the SETI.
SETI stands for the Search for Extraterrestrial Intelligence.
Right. And the idea is that if there are advanced civilizations out there, they could be sending us signals, either intentionally or unintentionally, and we could try to detect them.
And the radio frequency has some advantages because it is relatively transparent.
There's a range of frequencies. that will be very good for interstellar distance communication. visions of a young boy sitting in his bedroom late at night by the window and trying to tune an AM radio to get some far off.
Something from the middle of the country, say, if we're talking about California, but this is on a much, much broader scope, obviously.
When you're listening, what are you collecting?
I mean, you talked about the two-dimensional data, but what is the AI element letting you do?
Is it letting you sort of separate... signal from noise or kind of make better sense of what the data, you know, the radio signals as they are might be.
How are you leveraging the AI and machine learning?
Yeah, so primarily I work with the two-dimensional data, so the image data.
So in this data, we have a lot of noise, but also a lot of signals.
So we have these signals not because they're all ET signals, but because on Earth we make a lot of transmissions in a radio frequency.
And even though we build a telescope in radio quiet areas, we still collect a lot of these signals.
We call them radio frequency interference, RFI for short.
I'm just curious, do you know, is a lot of it these days from telecom, wireless phones, that kind of thing?
Yes, yes. So wireless phones and, you know, airplane transmissions, GPS satellites.
Right. So there's – and radars, you know.
So we see a lot of those signals. Some of them are more obvious.
Okay, that's obviously a radar, you know.
Some other ones are, or, you know, if we have a satellite catalog, okay, this satellite is moving overhead right now.
We see it right here. Right. But not all of them we can immediately identify to a known source.
So part of the job that the AI can do is to help us sort through the signals and try to characterize them.
We may not be able to know, okay, this is that type of radar, but we'd be able to know, okay, these signals all look similar, so they are probably from the same transmitter. have a sense are you basically looking for things that maybe there's there's more than one of and they look similar or But the distinguishing thing is they don't look like anything that you know.
Or in other words, you know, how do you know when you found something that that might be coming from, you know, Beyond Earth.
Right, right, right. So there are two ideas in here.
The field of SETI is fairly open-ended, right?
So we... It's just an idea that other advanced civilization might have developed similar technology as ours.
But in reality, we obviously don't know for sure.
So one idea is to, like you said, search for anomalous signals, signal that looks different from anything on earth.
And AI can certainly help with that. And during my talk, I discussed some of the techniques that can do that.
Another is a very, very primarily used technique to To find out if something is RFI is what I call spatial filtering.
So the main idea is that you observe multiple areas of the sky either at the same time or sequentially.
So these telescopes have very narrow field of vision.
So if you move... the pointing just a little bit, the telescope would not receive the same signal as where it was previously pointed. if you keep doing this and you keep on seeing the signal no matter where you're pointing the telescope, that's a good indication that that signal is actually leaking in from local.
Right. So that's the main observing principle in most SETI observations.
We're speaking with Jerry Jang. He's a graduate researcher at University of California, Berkeley, UC Berkeley, up in the Bay Area.
And he is working on... The search for extraterrestrial intelligence, looking for signs of advanced life forms out there in the universe beyond the Earth.
But let's go back a little bit and talk about your own backstory.
You've got a background in astrophysics as well as deep learning and data analysis.
How did you get into all this stuff? And how did you wind up in Berkeley doing what you're doing?
Absolutely. Thanks for asking. I have kind of more of a... roundabout story compared to perhaps a lot of young deep learning researchers nowadays.
We get all kinds on this show. Yeah. So I started out in theoretical physics.
I was, you know, when I was 12 living in China, Stephen Hawking visited China.
And I started looking into his book, A Brief History of Time.
And I was just fascinated with these big ideas that people have in trying to understand the world around us.
So after I read that book, I start learning more mathematics. and more theoretical physics.
So I studied theoretical physics during my undergraduate at Princeton University.
And one of the most fascinating thing to me about physics is universality.
It's not a strict use of a term because the word universality has specific meanings in physics.
But just loosely speaking, what I mean is that...
The study of theoretical physics is simple ideas, simple laws that govern a lot of phenomena.
So gravitation, for example, Newtonian gravitation, Anything that has mass has gravity.
It's all governed by this inverse square.
So that's, you know, the fact that the same mathematics, the same model applies to essentially everything is, you know, it's… the most mind-blowing thing to me.
And that's what I really loved about physics.
So as an undergraduate, I did work on particle physics.
And for graduate school, I moved to Berkeley.
I initially intended to work on cosmology.
And what's the distinction? What's cosmology?
Cosmology is study. So theoretical cosmology is study of the universe, how it evolves and evolves. models also you know fundamentally based on partially particle physics okay but also there's fluid dynamics there's there's There's obviously general relativity and a range of different physics.
Okay. So what I quickly realized that I did not quite enjoy doing cosmology as much.
I also quickly realized in astronomy, a lot of studies are about one very specific phenomenon.
And the excitement about this universal idea was missing to me.
Yes, right. Put it in a way, I was a little bit disappointed with my choice for a while.
Um, and then, uh, because Berkeley, you know, this was back in, uh, 2014, 2015, and, you know, TensorFlow was getting released.
Right, right. It was very exciting, and I started learning about deep learning and about machine learning. to my surprise in a bit, because I knew very little what to expect, that Actually, there is a lot of this almost universality in artificial intelligence because the same technique can be applied – from camera images to generating voice, to writing music, to finding aliens.
Yeah, yeah. I found that excitement again in AI.
That's fascinating. A couple of years later, I joined the Breakthrough Listen Lab, and they are amazing. looking for people to apply these modern techniques to this very exciting problem.
It's interesting the universality, you know, and that's kind of the driving force in your work journey here.
It's interesting to hear about. Also, I think it's a similar level of rigor.
So a lot of nowadays, It's less rigorous than mathematics, mathematics department, but it's more rigorous than say a lot of astronomy research.
So in that way, the two are also similar.
So kind of looking ahead a little bit, and whether in the short or longer term, what's driving you right now in your own work?
And then how do you see this evolution of AI, the compute power, but also now the The techniques and the software and the tools are so much more accessible.
How do you see that kind of shaping the search for extraterrestrial intelligence?
Yeah, absolutely. So personally, I'm driven by the universality of the method.
I'm actually not so much bound to the domain of SETI.
So I'm excited to see all of these new techniques.
The good thing about working in SETI is that there are relatively few people a small community, very few people applying these techniques.
So when new techniques come out, I get to try it first. on our data.
It's a nice perk. Yeah. So as I highlighted during my talk, one of the I think very exciting areas is semi-supervised learning.
So the radial steady data have the problem of being So we have large amounts of data, but we don't have the justifiable resources to have a human label them Like people do computer vision data sets.
And so the semi-supervised techniques basically combines the ability of these models to naturally synthesize, to characterize and study the data distribution and a little bit of nudge with some kind of label, maybe not so correct labels, because sometimes you can obtain labels that are Not so correct, but it still has some relevant information.
That way you can really teach the model to do certain tasks.
And in particular, I think... a good way to approach the problem is representation learning.
So representation learning are not specifically doing a certain task, but rather it's compressing the data.
So you have a model that compresses the data I'm talking specifically latent variable models, compress them into a lower dimensional representation. while maintaining some essential characteristics of the data.
So in doing so, you make the data a lot simpler to look at Because you suddenly reduce the original image down to essentially a single point.
And then when you have multiple examples of these single points, you can easily compare them.
You can see, okay, what it would be like to have something that's in between these two data points.
A lot of those things are possible once you have this latent representation.
Another good thing is that I think this way of formulating the problem help separate the machine learning from the astronomy.
Because the representation learning itself is a very well-defined machine learning problem.
And people work on essentially the same problem in different domains, in vision, in voice. so on and so forth.
Whereas once you have this representation model, the astronomers can take it because it's just a...
It's an operator that takes your data and compresses them.
It enables you to compare the different So they can take this as more or less a black box and design certain different kinds of pipelines to perform different kinds of tasks.
So by separating these two core modules of the problem, this would enable the machine learning experts to collaborate more easily with the astronomy experts.
So clearly anybody listening needs to go find, by the time you listen to this, I think they'll be available.
Go find the recording of your talk from GTC. and take a listen because I get the feeling we're just kind of scratching the surface of of some of this stuff here.
Jerry, if folks wanna find out more beyond your talk, if they wanna find out more about the work you're doing and the work that Breakthrough Listen is doing, Is there a website where can people go online?
Absolutely. So the primary website for the group is SETI.com. dot Berkeley dot edu but actually most of my work are not on there actually a lot of the even a lot of the work during my talk are are not yet published.
I see, okay. So once some of these works are published, we will put out some blog posts as well.
Perfect, cool. Stay tuned then everybody.
Jerry, thanks for taking the time to join the podcast.
And you know, it's, it's, It's fascinating in particular to hear you talk about the application to your domain, but then just...
And again, to use the word, the universality of how these tools can be applied to different domains and how... how that's one of the things that's really drawing you to the field.
It's very cool to hear you talk about. So thanks for joining us.
Thank you very much. Thanks for having me.
Thank you.
