Imagine if you had files equal to the number of stars in the sky. Some days, that might not feel realistic. Now imagine those files are about to explode. You can’t hope to save them all – but if you act quickly, you can save as much information as possible before it disappears.
Introduction
In the search for new exploding stars, astronomers are always in a race against time. “If something explodes, every day you wait to query that object deeper, you lose information because it’s constantly evolving,” explains Adam Miller, an assistant professor of physics and astronomy and the director of the data science fellowship program at Northwestern University.
What inspired you to build BTSbot?
We hoped to get a spectrum for everything that reaches this critical brightness, which teaches us about the rapid brightness survey. The way we can ensure we get a spectrum for everything bright is for someone to visually look at every reasonable filter, then decide “yes, we will try to get a spectrum” or “no, we won’t try.” These decisions come at a cost because we can’t take many spectra in one night. So it matters to make this decision well. The inspiration came from the fact that I was a bit tired of looking at the data and doing this repetitive task over and over every day. I thought – can we build a machine learning model that somewhat reduces the decision-making process?
What sets BTSbot apart from other telescopes using machine learning to search for explosive stars?
Many machine learning models that people build while looking at things like explosions operate in the realm of extracted data. Essentially, they measure features from images and then build models on those features rather than on the images themselves. We said, let’s try to solve this problem using the images. But the inspiration was to reduce the cognitive load on myself and on others looking at all this data.
How does BTSbot use machine learning?
With this specific task being performed manually on a repetitive basis, the idea is to integrate a machine learning model in this step of the workflow. I believe this is the most accurate way to apply machine learning to science. The alternative is to start with machine learning and then look for a place to apply it. I think it makes sense to start with a problem and see where machine learning fits into that gap. Now that we’ve actually begun running it, we sometimes see human operators still doing this manual check acting in strange ways. So we can tune the model to adapt to their habits, then it interacts better with humans. Also, once we started working in production, I’ve been monitoring it every morning to ensure it’s doing the right things. But the idea is to be a bit more to fully leverage the technology we’ve created.
How does your team at Northwestern collaborate with other researchers?
The overarching umbrella we all sit under now is the Zwicky Transient Facility project. This is an international collaboration among researchers at more than ten institutions. We all work with different but overlapping interests regarding the data generated by this project.
Do you use production AI in your workflow?
We have a webpage that collects and integrates our collective knowledge. Someone essentially created a plugin for ChatGPT on that specific webpage. For any new explosion we find, ChatGPT produces a three-sentence summary. Having AI that can generate these summaries and do so very quickly is really powerful and will be very useful. Especially since in a few years from now, we will have a new telescope that will increase the discovery rate of objects by several orders of magnitude.
Do you
Are there other benefits to using artificial intelligence in this work?
The robot telescope we use to obtain spectra is a small telescope. From a real financial standpoint, taking spectra with this telescope is inexpensive. If we create a machine learning model and it says, “these are the 10 most interesting things,” and it turns out that two of those things are completely uninteresting, that’s fine. We don’t see that as a huge negative cost. In fact, you could even say it’s a business cost. But the largest telescopes in the world – like the Keck telescopes in Hawaii or the Very Large Telescope (VLT) in Chile – these projects cost hundreds of millions of dollars to build. Their equivalent nighttime value is around $100,000. That’s the value of data from one night on one of these telescopes. If you’re able to look at eight things and mess up one, that means you’re wasting about $15,000. But what’s really beautiful about this work is that it lays the foundation for us to trust the system to make those very costly decisions in the future.
How do you envision your work being used by future researchers?
This is something Adam and I think about a lot – making it easier to launch new models and adapt existing models. One of the things I kept in mind while developing BTSBot is to make the code I write as flexible as possible. I want it to be very easy to take it and put it into the next task with minimal repeated effort. I want to use everything in common between this task and the next. I don’t want my work to go to waste. Similarly, because this code is general, it means others can also use it and adapt it to what they want to do. Making these tools public and well-documented makes it a little easier for people to join in.
What advice would you give to those who are curious but hesitant to automate their work?
I run a tutorial program for graduate students called the Data Science Fellowship program. The unofficial slogan for this program has somewhat become: care about the data. The risks in astronomy are not that AI doesn’t work, or that we should worry about whether AI is trustworthy or something like that. I think the biggest risk that can happen is that people don’t actually understand their training sets before they start running things. Many talk about doing something unbiased. That cannot be true in astronomy because someone at some point decided to point their telescope there. Any decision that led to pointing the telescope there instead of there is a bias. Maybe that’s not a meaningful bias depending on the science you’re trying to do. But I’m shouting at people, like, don’t worry too much about machine learning. At some point, you’ll put the data into some format. You’ll choose a machine learning model, then you’ll get some results and maybe choose another model and get slightly improved results. You’re more likely to run into problems if you’re not worried about the data, if you don’t understand your true starting point. Because I think that is incredibly critical in understanding what comes out on the other side.
What is the next task you would like to automate?
Finding these explosions as quickly as possible. It doesn’t necessarily save us time. But it gathers these spectra at times when humans typically wouldn’t get them. It opens new doors rather than saving time in the current ways. If these models could reach the point where we become more confident in asking more precise questions rather than just asking the question “will this become a bright star?” we can ask “is this explosion the result of a massive black hole being destroyed and now consuming a star?” If we could answer that, I would be very excited.
Done
Edit this interview slightly and summarize it for clarity.
Source: https://blog.dropbox.com/topics/work-culture/how-an-ai-powered-telescope-is-helping-astronomers
“`css
}@media screen and (max-width: 480px) {
.lwrp.link-whisper-related-posts{
}
.lwrp .lwrp-title{
}.lwrp .lwrp-description{
}
.lwrp .lwrp-list-multi-container{
flex-direction: column;
}
.lwrp .lwrp-list-multi-container ul.lwrp-list{
margin-top: 0px;
margin-bottom: 0px;
padding-top: 0px;
padding-bottom: 0px;
}
.lwrp .lwrp-list-double,
.lwrp .lwrp-list-triple{
width: 100%;
}
.lwrp .lwrp-list-row-container{
justify-content: initial;
flex-direction: column;
}
.lwrp .lwrp-list-row-container .lwrp-list-item{
width: 100%;
}
.lwrp .lwrp-list-item:not(.lwrp-no-posts-message-item){
“`
}
.lwrp .lwrp-list-item .lwrp-list-link .lwrp-list-link-title-text,
.lwrp .lwrp-list-item .lwrp-list-no-posts-message{
};
}
Leave a Reply