AI sheds new light on dinosaur footprints

Hey everyone, it's Tone here and today I'm really excited to be telling you about a project that's really close to my heart. A three-year collaboration with colleagues from Helmhold Center in Berlin, the University of Edinburgh, and Liverpool John Mo University on an engaging machine learning powered app. a virtual assistant that we hope will help researchers and students better understand dinosaur footprints. And it's all to do with their shape.

Introducing DinoTracker. 

Okay, so what is DinoTracker exactly?

DinoTracker is a powerful thinking tool that allows you to recognize the variation in shape of your very own dinosaur footprints and to test out specific hypotheses from identifying track makers to thinking about the environments that they once traversed.

Dinosaur footprints, I think, are some of the most fascinating, even the most iconic fossils, but they can be challenging to study because when you see a dinosaur footprint, the shape of that footprint is based on many different things. It's determined by the foot of the dinosaur, by the type of sand or mud that the dinosaur is walking through, and by the motion of the foot.

Using DinoTracker, this has helped us to understand the variation among nearly 2,000 dinosaur footprints. And that helps us have a more informed view of whether certain footprints were made by meati-eating dinosaurs or duck build dinosaurs. Or probably in the most troublesome case, there are tracks that are well more than 200 million years old, three toes, look a lot like the footprints of birds today. But if they are bird tracks, they would be many tens of millions of years older than any bird skeleton fossil. And that's why we need tools, better tools to understand the variation of dinosaur footprints.

I would say it's always super cool to have this interdisciplinary talk to each other effectively because all the involved persons were so nice to each other. It was really just fun, I would say. So the idea behind this is that we have the silhouettes of the footprints uh drawn by paleontologists and these images consist of thousands of pixels. So in images are high dimensional and what this disentangled variational autoenccoder tries to achieve is to really find find out what is varying independently in this tracks to find so to say the underlying core principle that created that footprint.

And for our application this core principle um is only eight dimensional, meaning only eight features are needed to describe these tracks, something like overall load or digit attachment. This is what is varying in that footprint and from this lower dimensional eight-dimensional space it then reconstructs the input footprint again. And based on these eight features, we can then compare and interpret each and every individual footprint.

So each of these eight features are in turn manipulable within Dino Tracker itself. These as you can see plot really nicely real time onto our exportable graphs opposite showcasing our footprint alongside a top result of seven further footprints from our data set with similar features.

But there is something else that we can do.

What we still need is a is an easy access visualization of all these 2,000 tracks that we have in our data pool. So you put in a new track and you want to see where is its neighbours, what is it close to? And TSN is a very popular technique from data science. And compared to other traditional methods, it tries to map this eight-dimensional space to a two-dimensional map. So to somehow roughly keep the distances in that eight-dimensional space on this two-dimensional space. Um, of course, it's not entirely possible, but it tries it as good as possible. And if you do this, you get that JS and E map which is available in Dino Tracker. And there you can just pull in your nuke footprint which you're currently analysing and it places this on a map alongside all of our data pool. So you can easily see if it's close to a theropot, a bird or an onto pot um or a tetropot.

For instance, when we look at some of these footprints that were made uh by some kind of dinosaur back in the triacic period, they look a lot like bird tracks. We see that actually the neural network model in the Dino Tracker app it it shows that they are most similar to bird tracks. It doesn't mean that they were necessarily made by birds but it means that yes what human researchers have recognized with their eyes as paleontologists when we look at these tracks and they look similar to birds that is true. The computer finds it as well. So either these tracks were made by birds that were tens of millions of years older than the previously oldest fossils or by dinosaurs that really really really back in the triacic period had feet that were similar to the feet of birds today or walked or moved in similar ways to birds today.

In another case, we use DinoTracker to look at some tracks from the Isle of Skye from here in Scotland where I work. tracks that that my team has studied and described over the years.

These tracks were made about 170 million years ago back in the Jurassic period. They are fairly large, about the size of a dinner plate, even a bit bigger, and they have three big toes.

Were they made by meat-eating dinosaurs? Were they made by duck build dinosaurs? If they were made by duck build dinosaurs, they would be perhaps the oldest members of that iconic group in the fossil record. DinoTracker tells us that some of these tracks from the Isle of Skye really are very similar to the tracks of known duck build dinosaurs. So we think it's very likely that some of these were made by pioneering duck build dinosaur species.
As a young researcher, I'm really grateful to my colleagues Steve Paige and of course lead app developer Gregor for giving me this opportunity to work on Dino Tracker, this innovative thinking tool that I hope you'll be able to draw on to better understand the variables that affect the shape of both dinosaur and bird footprints through feature comparison.

And the exciting thing about DinoTracker is that we're going to be continuing to develop it, adding more features that will allow you to make even better assertions on the footprints that you input into it. You can find D tracker on GitHub via the link in the description below. And if you have any thoughts or suggestions, we'd be very interested to hear them from you.

And yeah, thank you so much for watching.

I hope that you enjoy using DinoTracker as much as we did making it. And hopefully I'll see you in the next video.