Thanks to Brilliant for supporting  this episode of SciShow.

Go to Brilliant.org/SciShow   to check out Brilliant’s Introduction  to Neural Networks course! [♪ INTRO]. It’s tough to miss a big building-rattling  earthquake, but even if we do not feel a rumble beneath our feet, it may still contain  useful information for scientists.

That’s why 24 hours a day, seismographs are  listening closely to the rumbles of our planet. In fact, there are over 13,000  active seismic stations out there. And that is far more data than  seismologists have time to go through.

Which is why, in a 2020 study, researchers  set up a showdown of humans vs machines to sift through all this information and in the process, crown the heavyweight  earthquake detection champion. One type of rumble that researchers  are particularly interested in   is called a teleseismic event. That’s when the waves from a big  earthquake, often a long way away, trigger another, usually much smaller earthquake.

Geophysicists haven't come to a consensus  on how one earthquake triggers another in this way, so figuring out when it happens  and when it doesn’t may help us predict which areas are most at risk of  these remotely triggered rumbles. For the study, published in the  journal Frontiers in Earth Science, the researchers took data from  seismic stations around Alaska to detect both small earthquakes and  tremors caused by 30 larger quakes. Tremors are slower and weaker releases  of energy than earthquakes are, which release their energy near instantaneously.

Then, the researchers provided  the data to our competitors. First up, in the human corner: citizen scientists! Over 2000 “earthquake detectives”  volunteered for the study.

They were not experts in seismology, just  regular people who wanted to help out. They watched and listened to a  small piece of the seismic data, converted into both a visible  waveform and an audio file. The idea is that when we listen to sounds, our brains naturally do a great job of  analyzing and classifying what we’re hearing.

In this case, you can easily identify earthquakes  by a telltale sound similar to a slamming door. On the other hand, tremors sound like  a train traveling over railroad tracks, and background noise can sound  like anything from wind to static. So converting the initial data into a sound  made it easier to spot the differences.

Each of these pieces of data  was classified by 10 detectives as an earthquake, a tremor,  or just background noise. But they had competition. In the opposite corner was team 

AI: a machine learning algorithm. This was a deep learning model, meaning  it had the ability to classify data based on multiple features, as well as learn  from its past guesses and refine its predictions. It identified earthquakes by comparing  a section of unknown seismic data to a reference database where each  event was identified by experts. Essentially, the algorithm took a piece of  unknown data and found the most similar examples in that database to classify it.

To figure out which team came out on top,  the seismologist referees needed to identify some of the earthquakes themselves to figure  out what the correct answer actually was. And after all that, the winner is… team citizen science! Their earthquake labels were 85%  accurate, nearly 10% better than the AI.

They weren’t quite as good at classifying  tremors as they were earthquakes, but at the time there was no AI that  could detect tectonic tremors at all. So the humans still won out. Now, that does not mean the AI is out of the fight.

After this initial showdown,  things took an unexpected turn. The researchers used the citizen science  data to train the AI to do a better job. In another study published a year later, the AI was able to surpass the  humans at earthquake detection.

This training even taught the AI to detect  tectonic tremors for the first time ever. So maybe in the end, the moral of the story  is humans and machines accomplish more by cooperating than by competing. It’s likely AI will become an increasingly  valuable tool for finding earthquakes.

But the efforts of volunteer citizen  scientists will still be needed to first   point it in the right direction. This is the first time an AI has been trained to  detect this kind of remotely triggered earthquake. So in the paper, the researchers encouraged  other scientists to recruit volunteers to classify their data and grow the  size of the AI training database.

So humans will continue to help the robots  get better at detecting these rumbles, which in turn could help predict the risk  of future quakes and keep those humans safe! And while this particular project  is over, if you are interested in doing some citizen science yourself, we  will leave some links in the description. If you want to learn more about the  machine learning we talked about today, you can over on Brilliant, where they  have a whole course devoted to it.

They have plenty more science, math  and computer science courses too. And Brilliant has recently upped the  interactivity factor to the next level! Now, courses like algebra and algorithm  fundamentals are extra-engaging to guide you the whole way through.

If you’re interested, you can  sign up at brilliant.org/scishow to get 20% off an annual Premium subscription. So thank you for your support! [♪ OUTRO].