菜单

Machine learning in astrophysics

20 August, 2020 | 59 min 50 sec
Podcast Host Neal Ford and Rebecca Parsons | Podcast Guest Yogesh Wadadekar, Divya Oberoi, Shraddha Surana and Ujjaini Alam
Listen on these platforms

Brief Summary

Astronomers are increasingly turning to machine learning as a means to understand more about our universe — whether that’s the formation of galaxies or the Sun’s activity. Here, our co-hosts Neal Ford and Rebecca Parsons catch up with our special guest from the National Center for Radio Astrophysics in Pune, along with a couple of ThoughtWorkers to hear more about this intersection of data science and astrophysics.

Podcast transcript

Neal Ford:

Welcome, everyone, to the ThoughtWorks technology podcast. I'm one of your regular hosts, Neal Ford. And I'm joined today with one of our other regular hosts…


Rebecca Parsons:

Rebecca Parsons. I'm happy to be here. ThoughtWorks' Chief Technology Officer and one of your regular hosts as well.


Neal Ford:

And we have a large, varied group of people today, with a wide variety of recording qualities, because we're scattered literally all over the globe, and we're all trapped at home, because of the current geopolitical circumstances. But today, we're going to be talking about quite a fascinating subject, which is about machine learning and astronomy.


Neal Ford:

I want to get some folks to introduce themselves. We have a number of people here from the National Center for Radio Astrophysics, which is part of the Tata Institute of Fundamental Research based in Pune, India. And so let's get people to give an introduction to yourself.


Neal Ford:

I'm also going to ask our guests to specify whether they are talking about galaxies or the sun. Because we have two broad topics here.


Yogesh Wadadekar:

My name is Yogesh Wadadekar, and I'm an astronomer at the National Center for Radio Astrophysics, located in Pune, in India. Pune is a town with an urban population of about 6.5 million, and we are located about 200 kilometers southeast of Mumbai. So we are close to the western coast of India.


Yogesh Wadadekar:

We run and operate a very large radio telescope known as the Giant Metrewave Radio Telescope. This telescope is a radio interferometer, which is located about 80 kilometers north of Pune. Within our institution, we have researchers exploring a wide variety of areas in current astrophysical research. There's people who research pulsars, radio galaxies, they do cosmology with high redshift hydrogen detections and so on.


Yogesh Wadadekar:

So we cover a very large area, field of research in radio astrophysics, but we also do research, some of us also do research in fields of astronomy that are outside of the radio astronomy domain.


Rebecca Parsons:

And Divya, would you please introduce yourself?


Divya Oberoi:

Sure. Hi, everyone. My name is Divya Oberoi. I'm also one of the astronomers at the National Center for Radio Astrophysics. And like Yogesh said, we have a variety of different science areas in which we work. My own personal area of research is to do with the sun. I study the sun primarily at low radio wavelengths, which basically means that you're studying not so much the sun itself, but the very hot gas surrounding it, which we call the corona.


Shraddha Surana:

Hey, so I'm Shraddha Surana. I'm a lead data scientist at ThoughtWorks. Been with ThoughtWorks for quite a bit, but have been working specifically with the engineering for research team for more than a year. I think one of my first projects, I have been collaborating with the NCRA along with Dr. Yogesh and Dr. Divya in both their fields of specialty.


Ujjaini Alam:

Hi, I'm Ujjaini Alam. So I did my PhD in cosmology from the Inter-University Center for Astronomy and Astrophysics. And then for the last 15 years or so, I've been working in data science and machine learning algorithms as applied to cosmology.


Ujjaini Alam:

I joined ThoughtWorks last year as a data scientist. One of the first projects that I've taken up here is the radio solar project with Dr. Divya Oberoi.


Neal Ford:

Okay, great. So we have two scientists and two ThoughtWorkers, as our guests today to talk about two general problems in astronomy. So if we could have our specialists describe the two general problems that we're utilizing machine learning, to learn more about. So let's start with the galaxy’s problem, please.


Yogesh Wadadekar:

So what we are trying to understand is the star formation history of a galaxy. What that means is, whenever a galaxy forms over a period of time, it forms stars. Now, this star formation may happen all at once, or it could happen over a very long period of time. Or it could happen in bursts, which means there's star formation and it shuts off for a little while and it restarts again. And then there is a third burst or a fourth burst of star formation.


Yogesh Wadadekar:

So what we are trying to use machine learning is to predict what the star formation history of a given galaxy is. And how do we do this? We do this by measuring the fluxes of the galaxy, meaning how bright a particular galaxy is at different wavelengths or at different frequencies. And trying to use that to build a model of what the star formation history of a galaxy actually was.


Yogesh Wadadekar:

For different star formation histories, you expect to see different fluxes in the different wave bands. So that's exactly what we are trying to do.


Yogesh Wadadekar:

Traditionally, this is done using a very, very sophisticated technique called stellar population synthesis, wherein you actually try and model the star formation. You also try and model the kind of spectrum that stars of different masses and ages have, and how these stars evolve throughout their lifetimes.


Yogesh Wadadekar:

You also try to model how the spectrum of a galaxy gets modified by absorption and re-emission by dust. The dust itself could be very close to the star, in which case, it will be very hot. In some cases, the dust could be very far away from the stars as they form, and therefore, it's going to be quite cold.


Yogesh Wadadekar:

So, therefore, modeling this in a completely physical way, is an extremely difficult and complicated process. But this has been done very efficiently and very well, using all the known laws of physics and all the understanding of astrophysics that we have developed over the last century or so.


Yogesh Wadadekar:

What we've tried to do in our work is to try to see if one can take a physics free approach to this and just give a machine a very large number of examples, where we have measured the fluxes, and we've built, have the physics predict the star formation properties of these galaxies. And then we try to see whether the machine learning can build, without knowing the underlying physics at all, can build a model, which is equivalent to the physical model itself.


Yogesh Wadadekar:

Of course, here, we are trying to basically, the ultimate goal is to model the underlying physics based model and not the real universe. So, if the physics based model is an incorrect or incomplete representation of what we see, what the real universe is like, in that case, our model, our machine learning model will also be an incorrect and incomplete representation.


Yogesh Wadadekar:

But assuming that the physical model is a correct model of the underlying reality, what we try to do was trying to train a machine to understand, comprehend that model and make predictions. Of course, the physics based model is very compute intensive. It takes a fairly long time for it to model even a single galaxy, whereas a well trained machine learning model can do, one strain can make predictions in fractions of a second.


Neal Ford:

Can you give us a sense of what that data looks like? Is this all numeric data or is it imagery and the size we're talking about?


Yogesh Wadadekar:

Yes. So what we've done is, we've used data collected as part of a very large survey known as gamma. It's a survey that aims to understand galaxy assembly over different cosmic epochs.


Yogesh Wadadekar:

As part of the gamma survey, a sample of galaxies was first selected, for which spectroscopic observations were obtained with the Anglo-Australian Telescope in Australia. And following that, there were a lot of multi-wavelength observations that were obtained for the gamma fields.


Yogesh Wadadekar:

So for example, the gamma survey team went out and used an ultraviolet telescope to get ultraviolet fluxes for all the galaxies in their sample. Then they went and used infrared telescope to go to get infrared fluxes and then they used optical telescopes to get optical fluxes.


Yogesh Wadadekar:

Then what they did was they combined all the data that they had gathered on each galaxy and constructed a catalog. So in the catalog, basically, every line represents a galaxy and all the columns are numbers representing the fluxes in the various wave bands, going all the way from ultraviolet to the infrared, going through the optical wave band as well.


Yogesh Wadadekar:

So, what we are working with is a catalog of galaxies, that contains flux measurements in 21 different bands. This is not work that we've done, this is a very large community effort that was done over the last 10 years or so, wherein they obtained data on all of these galaxies.


Neal Ford:

And this is publicly domain data?


Yogesh Wadadekar:

Yes, this is absolutely public domain data. It has been released as a panchromatic catalog, wherein all the data have been combined. So they've done all the hard work of obtaining the data, processing the data and collating and collecting it together and making a public catalog.


Yogesh Wadadekar:

So all we had to do was to go to the gamma survey website, and simply download that catalog and that was our starting point.


Neal Ford:

Okay, I want to talk about the machine learning approach you took. But let's talk about the other problem first, which is the solar problem that we're applying machine learning to.


Divya Oberoi:

Okay, so let me try to put the solar problem in context. So it may seem like a bit of a surprise. But even though the sun is really the brightest object we see in the sky, there's plenty about it which we don't really understand yet. And a lot of that relates to the magnetic fields and the very hot gas which surrounds the sun.


Divya Oberoi:

Now, it has so happened that people have known for a long time that the sun as seen in the radio frequencies, especially at low radio frequencies, is much more dynamic, much more variable as compared to the boring sun which we see in optical. But we had never really had an instrument which was capable of being able to capture all that detail.


Divya Oberoi:

So sometime around 2013, a new instrument became available in Western Australia, it was called the Murchison Widefield Array. And this instrument is the best we have on the planet for capturing in great detail how things are changing on the sun, over very short intervals of times, over a fraction of a second. And over very small spectral spans, you move 100 kilohertz in the frequency which you're observing, and the sun looks completely different from what it was at the neighboring frequency.


Divya Oberoi:

So we had been working on building a robust, unsupervised pipeline, which would make these images in an automated manner. And just about the time ThoughtWorks approached us for this, we had basically come up with a fairly good pipeline which would deliver these images. Now, we can in principle make something like 100,000 images a minute.


Divya Oberoi:

Having solved the problem of imaging, we immediately came across the next problem, which is how to understand what all these images are telling us. It was really impossible to look at them image by image and to imbibe for the human brain to wrap one's head around what all these thousands or hundreds of thousands of images are telling us.


Divya Oberoi:

So we wanted to, or started to look for a machine learning based approach, which would help us synthesize the information content of all these very large number of images. That's how we got started here.


Neal Ford:

Okay, great. So let's talk, now let's go back to the galaxy problem and talk about the data science machine learning aspect of that problem and a dataset that looks like that.


Shraddha Surana:

So for the star formation histories problem that we were trying to solve, the data that we used for it was from the gamma survey. So, this is a pretty standard data and the kind of data that we would use for most machine learning models.


Shraddha Surana:

So, there are some input records, which are the flux values in different frequency bands. But along with that, we also have the output free parameters that we are trying to estimate over here, which is namely the star formation rate, the dust luminosity and the stellar mass of the galaxies. So, these have already been estimated by the MACFES model and we have those values.


Shraddha Surana:

So the idea over here was to implement a machine learning model that will be able to mimic the current MACFES model that is present. We used several machine learning algorithms to see and the one that we zeroed down on was deep learning. We got pretty good accuracy for the deep learning model that we implemented and we created three models to predict each of those parameters.


Shraddha Surana:

Now, the main advantage that we had with this was that while the current MACFES model tends to take more of a brute force approach, in the sense that it has millions of templates that it tries to fit through each reading of the galaxy, which takes about 10 minutes per galaxy. Now, once we train a deep learning model for each of these parameters, and the training takes a maximum of 30 minutes for one of the parameters, we can now predict the same parameters for millions of galaxies in just a couple of seconds. So there is a huge savings in terms of time, which will help expedite the research project and the inferences that can be made from these estimations.


Shraddha Surana:

So one of the challenges that we had was to understand why most of the results, so we got pretty good accuracy, the estimates that the deep learning model was predicting, were pretty close to the MACFES model. However, there were some results that were not very similar, that had a stark difference. Now, machine learning models tend to generalize the results or at least that is the way that we try to implement the machine learning model.


Shraddha Surana:

For example, take ohm's law. So if the readings are taken of the resistance, current and voltage and we create a machine learning model, a regression model, it will essentially be able to give us that particular formula. It will generalize, even though some of the readings may not match or adhere completely to the formula, it is a similar thing that we are trying to do over here with the deep learning model, as well that we have implemented.


Shraddha Surana:

So, in such cases where there is a stark difference between the deep learning estimates and the MACFES model estimates, these give us very interesting cases to actually go further and investigate as to why the difference is. Is it something that can be improved in the deep learning model? Or can it enhance our assumptions and modeling of the MACFES model, which in turn might lead us better understanding of these galaxies?


Shraddha Surana:

Yogesh, do you want to add anything?


Yogesh Wadadekar:

Yes. I'd like to add a couple of points to what Shraddha has just said. First of all, I'd like to say how and why we chose the output parameters. So what we are trying to do here is to use machine learning to predict three numbers that characterize the star formation history of the galaxy.


Yogesh Wadadekar:

The first number is the stellar mass, which represents how much mass is contained in the stars that have formed within the galaxy. Obviously, if the star, if the galaxy has formed a lot of stars over its history, the stellar mass will be high. And therefore, the stellar mass is a measure of the integrated star formation history of the galaxy.


Yogesh Wadadekar:

The second number that we'd like to predict is, what is the current star formation rate? Which means how many stars are being formed in this galaxy at the current time? Now, this is completely different, because it's unlike the stellar mass. This is not an integrated quantity. It's a differential quantity, we are trying to measure how many stars are being formed per year in a particular galaxy.


Yogesh Wadadekar:

The third parameter that we are trying to predict is the dust luminosity of each galaxy. The dust luminosity is a number, which depends on how much star formation has happened in the galaxy. Because if you go back before to a time when no stars were formed, there is little or no dust in the galaxy. Galaxies increase the quantity of dust within them. Because dust gets formed, dust which is basically silicates, get formed inside of stars. So it's only when stars evolve and the first generations of stars return their materials to the interstellar medium, do we get dust within galaxies.


Yogesh Wadadekar:

And the dust luminosity naturally depends both on how much dust there is, and how hot it is. If there is a lot of dust and if it's at a high temperature, then the dust luminosity tends to be very high.


Yogesh Wadadekar:

So these are three independent quantities that together helps us understand how and when stars form within a particular galaxy.


Neal Ford:

You're basically saying the more crowded the house is the dustier it is.


Yogesh Wadadekar:

Yes, in some sense, yes. If there are a lot of stars that have been formed already, then it does get more dust here, because those stars evolve, some of them may explode as supernovae, and then they throw out a lot of material into the space between the stars. And then the dust begins to increase.


Yogesh Wadadekar:

The second point I wanted to make was regarding the extremely non-linear dependence between the fluxes that we measure and these output parameters. Unlike the Ohm's law example, that Shraddha mentioned, which is a linear relationship between voltage and current, this one is a very non-linear relationship. So, tweaking one parameter by a small amount can change the output by a very large amount.


Neal Ford:

So this is sensitivity to initial conditions.


Yogesh Wadadekar:

There's a fairly large sensitivity to initial conditions. So it's very important to measure these fluxes as accurately as possible.


Neal Ford:

This is a problem where you decided to use a technique called supervised machine learning, correct? Could you describe what that is and why that approach was chosen?


Shraddha Surana:

Yes. So in this particular case, so actually, broadly speaking, machine learning techniques can be classified into supervised techniques and unsupervised techniques. Supervised techniques is when we have a data where the inputs are matched to outputs.


Shraddha Surana:

So to take a simpler example of common house prices, based on certain factors, such as the number of bedrooms, the carpet area, et cetera, and the price of the house is available to us, that's when we would go for a supervised machine learning approach, where we are telling the model that these are the inputs, and this is the output for it. And hence, in this particular case, where we have to predict these three star formation properties, it was clear from the beginning that we have to use, that supervised learning techniques would be most suitable.


Shraddha Surana:

Because we had the input flux values and we had the corresponding free parameter estimates given by the MACFES models.


Neal Ford:

Okay, so let's talk about the solution to the solar problem. And it actually utilizes a different kind of machine learning. So if you could describe the approach there and how that approach was different than the galaxy problem.


Ujjaini Alam:

Of what outcome we desired, we had all this data, we knew that there must be a lot of physical information in it. But we didn't know what question we were asking. Therefore, in this case, unsupervised learning seemed to be the obvious way to go.


Ujjaini Alam:

And that's when you have a machine learning algorithm, which is trying to draw inferences from a dataset, which has input data, but two labeled responses. The commonest way of doing this is to do some sort of clustering, which means it takes all the data and it tries to find patterns, hidden patterns of groups in the data. And then the expectation is, once you find those patterns, you should be able to look at those patterns and draw some inference.


Ujjaini Alam:

So in this case we had, I'll take a minute to describe the data, again. What we started with was a four minute data cube, which had data every half a second in time and 48 frequency bands. And this was over a 200 by 200 pixel image.


Ujjaini Alam:

So we decided to start off with treating each image as a data point. So we have 22,000 data points over different frequencies and different time slices. Each data point then contained 40,000 pixel intensities and these were the features of the data points. And then we tried to use unsupervised clustering algorithms on these data points.


Ujjaini Alam:

What we found was, firstly, that when you have such a large number of features in the data, that in itself causes a problem. Most unsupervised learning algorithms do not scale up to this kind of large number of dimensions. We used self-organizing maps, we used density based clustering algorithms, and in each case, we found that having so many features was a problem.


Ujjaini Alam:

So we had to take two approaches to that. One was to do some sort of feature selection, find out the feature importances. The other was a more data engineering approach, which was to figure out if we could speed up the processes by using say a different language, C++ instead of Python, or if we used GPUs, whether this would improve the results, speed of the results, et cetera.


Ujjaini Alam:

But then the problem with this was, since the pixels are the features, if you do feature importance, and if it tells you that say pixel number 125 is important to the clustering, that does not necessarily lend itself to a physical interpretation. Because pixel number 125, you're not very clear what that means. So we did have problems. We got very consistent results from various clustering algorithms.


Ujjaini Alam:

But once we had these results, we could not really, physically interpret those results. We couldn't quite figure out why those clusters had formed. We knew that they were forming over and over again, with different methods, but we didn't know what was causing them to form. So then we had to take a step back, and we had to think of how to represent the data in a better way.


Ujjaini Alam:

What we decided was to concentrate on the peaks within the images. So rather than considering the entire image, we considered the peaks in each image. We tried to fit those with Gaussians. We tried to see what information those peaks could give us. And we found that if we took the peaks and followed their time series, and then we tried to cluster those time series, we were getting better results.


Ujjaini Alam:

So, in the end, we ended up using clustering, but in a very different way than we had initially envisaged. And that is something that I think happens a lot with unsupervised learning, because you don't know what question you are asking. You have to often step back and modify your question and keep modifying your methods, based on the results that you get.


Ujjaini Alam:

So this is a work in progress. At the moment where we are is that we are trying to cluster the various peaks in the images into different classes. And we are hoping by doing this, to isolate a specific type of physical process, which are the nano flares. And Divya we would be the better person to talk about what nano flares are. Why are they important in physics?


Neal Ford:

Before we go there, a couple of questions for you. So, there's a famous quote by the artist Picasso that said computers are useless. They can only give you answers, but not questions. But it sounds like you figured out a way to get a computer to give you questions as well, which is really, really useful.


Neal Ford:

And the other thing and this is a broad generalization that I want to, that's translated into layman's terms, and I want to make sure that I'm understanding the difference between these two. So is it fair to say that you would typically use supervised learning if you have a known, unknown problem, and use unsupervised learning if you have an unknown, unknown problem?


Shraddha Surana:

Just to add to that, for our unknown, unknown problem, we would typically go for unsupervised machine learning problems. But you could also use it for known, unknowns as well or unknown, knowns, known, unknowns rather. Because at the end, what unsupervised machine learning algorithms do is to figure out groupings, clusterings and patterns that are present in the data and that can be useful in both the scenarios.


Rebecca Parsons:

It's definitely applicable in many places. We've been using methods like this literally for decades. For example, credit card fraud detection, if you have a particular spending pattern and you aggregate those spending patterns, and then when some anomalous spending arises, it gets flagged. So there are numerous applications for this outside of the world of science and in terms of price predictions, recommendations, et cetera.


Rebecca Parsons:

But I want to get back to the sun. So, Divya, what can you tell me about some of the impacts on your way of thinking about the sun have arisen from some of the results that have come so far?


Divya Oberoi:

Before I go there, let me make some comments about the solar app, the machine learning applications to the solar work. And something which I did not mention before is that not only we have an enormous number of images, individually, these images are very deep. Their dynamic ranges can be as high as 100,000. So the contrast between the brightest and the faintest reliably detected feature could be as high as 100,000.


Divya Oberoi:

And that is about two or sometimes even three orders of magnitude better than what has ever gone before. So that is an enormous part of the phase space, which has never been explored before. We also know that the sun hides many secrets, in terms of what Ujjaini had mentioned as nano flares, which is a very large number of very small events, like small smatterings, or a large number of individual fireworks going on, on the night of Fourth of July. So many, so that you can hear only the rumble and you can't make out the individual firework.


Divya Oberoi:

So, here also, on the surface of the sun, there is this general idea that there is a very large number of individually very small events, which are taking place all the time. And they have just been beneath our sensitivity limit for thus far.


Divya Oberoi:

And there has been a regular attempt or a long standing attempt to go to fainter and fainter events, more and more sensitive instruments, which would allow us to detect them. And now for the first time, we have evidence for having detected some of these in the solar data. And so those tiny Gaussians which Ujjaini was referring to, we believe those are coming from the sites where these very tiny explosions, if you will, are happening and we are just trying to characterize them, to see if they are consistent with our expectations based on the theory.


Divya Oberoi:

And then to get back to the question which Rebecca had asked, that how has this changed or influenced my perspective about solar physics in some ways, I think what it is telling me is that, well, okay, when we started down this path, another reason for me to go for unsupervised learning was to make sure that we were not biasing what we were looking for, just based on our prior knowledge. Because we are in completely uncharted waters, in terms of how faint features can we see. I wanted an algorithm to find for me whatever relationships exist in these data, rather than me directing the algorithm to look for specific relations, which I know must exist.


Divya Oberoi:

In the process of looking at these data, what we have ourselves discovered, not quite using machine learning, but just by staring at these images long enough that there are many, many faint relationships which exist. Relationships between the size of something and its brightness, which vary in some sort of a quasi periodic manner. Things with periodicity is ranging from a few seconds to a few minutes, none of which we knew about earlier.


Divya Oberoi:

Now, we happened to find them, because we happened to be staring at some particular four minutes of data for a long enough time and noticed it. But what it tells me now is that these data are really rich, in terms of what must exist in these data, the various correlated variations, in terms of what is happening to the morphology of that emission, to where it is located, to maybe a small jitter in its motion and to the variation of its amplitude.


Divya Oberoi:

There are good reasons to believe that there is a large number of relationships which exist between various parameters in the emission, which is sitting in these images. And that is what I would love these machine learning algorithms to discover for us.


Rebecca Parsons:

Great, thank you. So, let's talk a little bit about how in both of these instances, you address more unexpected results. I know Ujjaini talked a little bit about how the focus was shifted from all of the data to the peaks. Shraddha was actually talking about some of the ways that we're learning something when the machine learning predictions seem to be off from the model.


Rebecca Parsons:

I wonder if you can talk a little bit about what kind of feedback loop you see between the insights you're getting from the models, and what that is telling you about the physical system. Professor Divya, you were going there a little bit towards the end. But Dr. Yogesh, could you talk a little bit about what we are learning from the perspective of our understanding of the underlying physics of galaxy formation? Has that work been informed by some of the things that you've discovered as you were working to try to increase the effectiveness of the model?


Yogesh Wadadekar:

Yes. So for me, one of the most interesting things are as you pointed out, the failure cases where the machine learning model made a prediction that was way off from what the physics based stellar populations synthesis model gave us. We have only a handful of such objects, wherein the prediction is way off. But the most surprising thing which I found was, when we used the machine learning model, what I had expected was that when you made a joint prediction of stellar mass, star formation rate and dust luminosity together all in one go, one should get a smaller scatter than if you tried to predict each one individually.


Yogesh Wadadekar:

The reason for that is that these are not completely independent parameters. They're coupled to each other in complex ways. So if you try to predict one without trying to predict the other simultaneously, you should get a higher scatter.


Yogesh Wadadekar:

But surprisingly, what we found was that if you predicted each of these parameters separately, the scatter was actually lower than when you tried to do them all together. And this was significantly lower in a statistical sense. So this is something that I still don't understand. That's something I would like to investigate further, why this is happening.


Yogesh Wadadekar:

The other things that I would like to investigate are these outliers and whether they can help us inform the physics. So there are clearly objects where the prediction from the machine learning model doesn't match up with the physical prediction. And we've tried to take a quick look at these objects, but we haven't found anything completely different in these objects.


Yogesh Wadadekar:

So this is something that is very interesting, needs to be done. But we were focused on writing up our work as a research paper and getting it accepted and so on. So we haven't yet looked at these objects, but these are potentially the most interesting.


Yogesh Wadadekar:

Even if it turns out that we cannot understand why there is such a discrepancy, these objects are potentially very interesting. Because they're fundamentally different from the average population in that region of parameter space. So what we've done is, the machine learning tries to get things right on the average, and therefore, it could miss out unusual objects of which it has not seen an example of in its training set.


Yogesh Wadadekar:

So it's very likely that these objects that are outliers in the predictions are actually very unique galaxies whose star formation properties are very different from other galaxies of similar mass and similar star formation properties. So we'd have to look at these objects individually. We haven't done that yet, but that is on our immediate agenda.


Neal Ford:

So could we get the two scientists to remark on how much machine learning and computation in general has changed your fields over the last few years, in the last decade or so?


Yogesh Wadadekar:

I can comment on that. In astronomy, as you know, we study the whole universe. When we study the whole universe, there's a lot of objects to study because it's a very vast universe.


Yogesh Wadadekar:

Also, over the last 80 years or so, new windows on the universe have opened up. So radio astronomy started in the 1930s and '40s. Later on, we had X-ray astronomy, and gamma ray astronomy, and so on. So different kinds of telescopes could be built. Some of them needed to be located in space. And they gathered data, the amount of data that we gathered began to increase.


Yogesh Wadadekar:

So we had innovations in a telescope technology, but that was matched with the continuous growth in the capabilities of computers, in the amount of data that we could gather, the amount of data that we could store, the amount of data that we could process, and so on. Everything has been growing exponentially over the last, not just the last decade, but over many decades.


Yogesh Wadadekar:

The kind of work that we are doing now, as was already pointed out by Divya, depends on data that could not have been gathered in the past. So Divya talked about the MWA, which is a very unique telescope. The kind of data that it gathers was not available a decade ago.


Yogesh Wadadekar:

Similarly, the kind of large catalogs of galaxies on which my work with ThoughtWorks depended, requires you to have telescopes that are capable of carrying out large area surveys in different wave bands. And this is a capability that we've only had over the last decade or two. And therefore, this is something that could not have been done before.


Yogesh Wadadekar:

And of course, with the large availability of data, we are facing a very unique constraint in the past, as compared to the past. In the past, when I started my astronomy career 25 years ago, we actually put in a lot of effort to gather data for our research. Now, because of the explosion of technology, both the computer technology and the telescope technology, the data are there for the asking.


Yogesh Wadadekar:

So when we started working on this research project, the gathering of the data, which used to be an effort, which used to take years, is now a five minute job. Basically, you head off to the website, you figure out what dataset you need to download, and you download it.


Yogesh Wadadekar:

Another thing that I wanted to mention, which is unique to astronomy, is that astronomers are possibly the only scientists who have very large amounts of data and are willing and able to share it openly with everyone else. In many other domains, for example, in medical studies and so on, for various reasons, including that there is a lot of money to be made in that profession, it's very hard to get access to data without signing various non-disclosure agreements and so on.


Yogesh Wadadekar:

In astronomy, it's like, oh, you want data? Here's the website. So astronomers not only have a lot of data, they're willing to share it with people. And they go to great lengths to make that data available in a form that people can simply download and use.


Neal Ford:

Okay, we're near about the time limit of what we normally have for length. Does anybody want any last questions or last observations?


Rebecca Parsons:

I just saw Dr. Divya come off of mute and then whoever else was just saying that, so maybe we can give them each a chance to weigh in.


Divya Oberoi:

Sure. Thanks, Rebecca. So what I wanted to say was that the way I got started on machine learning or looking at machine learning for answers to this was that I was working with a student. We picked up a fairly run of the mill kind of an event which takes place on the sun and said that, okay, we'll put it under a microscope. And by under a microscope, I really mean put it, observe it using our telescope in such excruciating detail that nobody has ever done it before.


Divya Oberoi:

And what we found there surprised us, in the sense that we found relationships between various aspects of this burst, which we never knew existed. And then as we were working on it, we found that, hey, there is something else which is going on, not so far away from this one. Let's look at that as well.


Divya Oberoi:

And when we did that, we found a completely different zoo of animals there, where again, there existed relationships, which were very different from what we found in the original event. And the process of really figuring all of this out and understanding them took us better a part of two years. And all of this is happening in just four minutes of data.


Divya Oberoi:

So it really struck me that there is so much richness in these data. So many things are happening in these data, that it is probably going to be very hard for the human mind to appreciate all of them, to wrap one's head around all of them, and then figure out what these are telling us.


Divya Oberoi:

So it became imperative to me to look for things from the machine learning, AI sort of world to at least help us identify what relationships exist. And then we use our existing knowledge of physics to figure out what it is telling us about the sun.


Shraddha Surana:

I had two points that I wanted to highlight. Having worked in some of the business problem statements earlier and now working specifically for research.


Shraddha Surana:

I felt there were two things that were starkly different. I think the first that I want to highlight is the explainability and the reasonability of the problem statement. I think this is very important when we are approaching research problem statements, because the idea in this particular case is not only to build a model to do something well, but also be able to understand what is happening behind the scenes.


Shraddha Surana:

So, for example, understand the relation between the input values and the output values like in star formation histories and when they are different from the current model, to really be able to understand and explain why that is. This is very different from some of the business problem statements where, for example, if there is a phase detection problem statement.


Shraddha Surana:

In those cases, we are mostly interested in optimizing the model, and not very much bothered with why the model is behaving the way it is. So this is one thing that I wanted to highlight. And of course, this is also important in the unsupervised learning problem statement that we are doing for the solar images, which is to be able to explain what is happening in the data. Needless to say, I feel once we master this in the research aspect, it has direct usefulness for any business problem statements as well.


Shraddha Surana:

The second thing that I wanted to highlight was the vagueness of the problem statement. So, for example, for the radio solar imaging problem that we have, which is an unknown, unknown problem statement, and it's quite vague actually. So, which is very different from most business problem statements that I have worked with, which was, for example, optimize my product prices or create chatbots or identify customer churns, which have a concrete problem statement or at least a direction that I want to be able to optimize my supply chain or something.


Shraddha Surana:

So this is very different, which is here is data. We know from past experience that this sort of data has something useful in it. We want to be able to bring out that usefulness faster than usual. Because many humans looking at it takes that much more time and that much more humans. And that is where we want to be able to bring in machine learning models, which can at least dig into the data and give some direction where we can explore further. Explore with other statistical methods or with machine learning itself.


Shraddha Surana:

It's more like finding a needle in the haystack, and I'm pretty optimized that we will be able to find something. And that's where the excitement lies for me. So these are the two things I wanted to highlight, the explainability and reasonability that is required and the vagueness that we have to deal with.


Shraddha Surana:

And I think this also has direct business implications, because I do hear colleagues mentioning that we have clients saying that this is the data, figure something out. I think if you master that here, we'll be able to implement it in many places.


Yogesh Wadadekar:

I just wanted to say that astronomers have now started to look at machine learning seriously, not because they like machine learning, but because they simply have no choice. The amount of data that we are generating from the present generation telescopes, as well as the data that we will generate over the last decade is so huge, that it's simply impossible for a human to delve into these data and gather all the scientific insights. There's simply no choice but for us to automate these things in various ways.


Yogesh Wadadekar:

And machine learning is one of the ways in which we can look for the most interesting, physically most interesting objects in the vast datasets that we collect. So unlike many other fields, we are not just interested in the needles in the haystack, but we are also interested in the haystack as a whole. We're interested in doing statistics with all the objects that we observe. We're also interested in individual, unusual objects that we never looked at before.


Neal Ford:

It seems like astronomy is the latest scientific discipline who's realized that they eventually become a data driven discipline, at least in part. Biology went through that and your observation about the amount of data that's available, essentially forces astronomy to start becoming a data intensive and a compute intensive field, at least partially.


Yogesh Wadadekar:

Yes. It's even more severe in astronomy, because in fields like biology, the data are growing fast. But the number of people looking at the data is also growing very rapidly. The number of astronomers in the world is relatively stagnant.


Rebecca Parsons:

Well, hopefully we can change that and get more people interested in astronomy, or at least computational astronomy.


Yogesh Wadadekar:

Absolutely. Sure, sure. Because astronomy is such an open field, there is a lot of room for people who are not professional astronomers to come and look at astronomical data and look for things that we are not geared to look for, especially for the unknown, unknowns, there's a lot of room to make serendipitous discoveries that us professional astronomers will not make because they're simply not looking... They know what they're looking for and they're only looking in a particular direction, in a particular way. But somebody who's completely unbiased, might be able to make discoveries that would have been missed by professional astronomers.


Neal Ford:

So instead of burgeoning astronomers making their own telescope at home, maybe they should be learning machine learning and algorithms and making scientific discoveries in astronomy that way.


Yogesh Wadadekar:

Absolutely, absolutely. And this is something now that can be done by anyone, sitting anywhere. You don't have to worry about the weather, because so long as your internet connection is up, there's a lot of science that can be done by mature data scientists in the field of astronomy.


Rebecca Parsons:

So in the chat, Prasana has raised the question of data engineering, so Shraddha, Ujjaini, is there anything in particular from a data engineering perspective? Yogesh just talked a great deal about the sheer volume of data. But are there things just from an engineering perspective and dealing with the data?


Rebecca Parsons:

I know Divya, you referenced that when you were talking about focusing initially on all of the data, and then just looking at the peaks. Anything interesting from a data engineering perspective that we should mention?


Ujjaini Alam:

I talked a little bit about this when I talked about the problem. So when you have a large amount of data, and a large amount of features in that data, there's two ways to deal with that problem. One is from a data science perspective, you try to reduce the number of features, you try to just look at only part of the information that is important. Feature selection, feature importance, calculation, those kinds of things.


Ujjaini Alam:

But there is also, as your dataset becomes larger and larger, it becomes a data engineering problem. At the moment, we are looking at four minutes of data for a particular dataset. The full dataset is 70 minutes, which we have not looked at as of now. But if and when we have to look at that 70 minutes of data, and we have to apply the algorithms we are applying right now, then we have to do some major refactoring to our algorithms. And that would be from a data engineering perspective, where you look at a different language to speed up the process, when you look at GPUs.


Ujjaini Alam:

So, that is something we have partially worked on. We wrote a self-organizing map algorithm in C++ in order to speed it up when we initially had problems. It is something that as the dataset gets bigger and bigger, we have to deal with it from a data engineering perspective, not just a data science perspective. We have to think of it in terms of creating pipelines to deal with these huge amounts of data, rather than just the data science part.


Shraddha Surana:

For the star formation histories problem statement, although the data did have lacks of records, we didn't have a requirement of having a computation intensive processor or having any pipelines. So this can be treated as like a total data science problem.


Yogesh Wadadekar:

One more thing I wanted to add, these two problems, I believe, are just the beginning. When we started working with ThoughtWorks, we astronomers at NCRA came up with about a dozen different problems where a machine learning approach could be applied in the astronomy domain. And we asked our partners in ThoughtWorks to select one or two of these problems to get started with, and they happened to choose these two problems.


Yogesh Wadadekar:

But we have many more problems on which such an industry academic partnership could be usefully employed, in order to tackle those.


Neal Ford:

Okay, well, I want to thank all of our guests today, both ThoughtWorks and our scientists for giving us this great insight into machine learning, and how it is being utilized in the real world, including a great example I think of the difference between supervised and supervised machine learning.


Neal Ford:

I wanted to mention ThoughtWorks Engineering for Research is organizing 2020 symposium on the theme of the role of artificial intelligence in scientific discovery, will AI ever win its own Nobel prize? That will be in Pune, India. It was originally scheduled for June of 2020, but is going to be rescheduled for a date yet to be announced. So, keep an eye out for that. That should be fascinating.


Neal Ford:

So thanks very much for our guests and for Rebecca.


Rebecca Parsons:

I would like to add my thanks to Dr. Yogesh, Dr. Divya, and Shraddha and Ujjaini from ThoughtWorks for this fascinating discussion.


Check out the latest edition of the Technology Radar