Thank you to Cometeer for  sponsoring today’s episode!   Cometeer offers barista-quality  coffee that’s delivered in frozen,   recyclable aluminum capsules.

Click the link in the description and you’ll get $20 off your first purchase, plus free shipping. That’s 10 free cups of coffee and   over 30% off! [♪ INTRO] The art of writing is  thousands of years old, and, amazingly, some texts have even lasted that long.

Whether etched on paper or stone, the bits that   survive give us lots of insight into ancient worlds and their cultures,   but not before going through some hoops. You see, some of these scripts are divided across   thousands of tiny fragments, which are tricky to read,   even by skilled researchers. So, experts spent more time   figuring out what something says rather than understanding what it really means.

But computer vision is starting to change that. Reading fragments of documents to decipher individual letters and words  is pretty challenging. So, the main goal for historians is to  leave that tedious task to the machines.

Computer vision can process images  of text with algorithms and artificial intelligence to automatically  extract the text from those pictures, producing a clean digital text that experts  can easily study and share with others. While computer vision has  been around in many forms, today's most powerful techniques  tend to rely on deep learning. Deep learning involves a unique  algorithm called a neural network, named for its “network” of interconnected  nodes, which are called “neurons.” The neurons, in this case, are  simple mathematical functions.

They're arranged in a sequence of layers,  so that when data are fed into them, the data are processed from one layer to the next to make a prediction or classification. As the name implies, neural networks are sort of analogous to our brain structure,   albeit very simplified. So, modern networks   have multiple layers of neurons, which scientists describe as “deep.” And they become helpful at  certain tasks by learning from different data examples,  hence the term “deep learning.” Now the idea is that scientists repeatedly show the network pairs of inputs and outputs.

The inputs are data the  algorithm needs to process, like a digital image of a text fragment. Outputs are the desired outcome or what   the text from the picture is supposed to say. So just to simplify, researchers might show the   network an image of the letter “R” and let it make a prediction   about which letter it is.

It might guess correctly, or it   might guess wrongly and classify it as a “Q.” When the algorithm classifies something wrong, that's when the “learning” comes in. It involves using the difference between   the desired outcome and the algorithm’s guess   to tweak the network’s parameters. So it can then make a more accurate   prediction the next time around.

By showing thousands of these kinds   of examples, again and again, the parameters of the network   configure themselves to get better and better   at figuring out what the images say. With enough data, networks can eventually   correctly identify letters and words they have never even seen before, which is how   we check how accurate they really are. Neural networks are a pretty general tool, so it can do more things  beyond just reading images.

They can be trained to identify which parts  of an image actually contain the text in the first place, make an attempt to  fill in missing text, and even digitally reconstruct whole documents that are  too fragile to physically touch. Most of these applications currently  involve using lots of training data of   correctly labeled images with the correct  output, and the more, the better. But the correct output, in this case, usually comes from humans who read   the text in the first place.

So, in this kind of learning,   machines might come to basically copy the examples humans give them. And, as you probably know, we sometimes  make biased or flawed judgements, so machines can come to inherit  the same kinds of biases. Thankfully, in the case of text, the output  we want tends to be pretty unambiguous.

So, for now, it still takes lots  of initial human work to create enough data examples for these algorithms  to “learn” how to do their jobs. But once you go through that initial  effort, scientists can train models that work quickly and efficiently  on ancient documents. So far, computer vision has been  used to read printed Latin letters, like the kind used for English, for  decades, but new techniques are expanding to different historical languages and  writing styles from all over the world.

Researchers have created models that can  read stylized versions of Latin script from old German, Tamil, Devanagari, the  ancient Ethiopian language of Ge’ez, Korean, and Japanese, just to name a few. And, as we mentioned, these algorithms can   do more than just read text. One 2021 study used computer   vision to digitally piece together torn papyrus fragments from the Dead Sea.

Another 2021 study made old documents  with faded text easier to read, while a different study from 2019 helped  guess missing ancient Greek words in broken stone tablets  based on existing examples. These algorithms might also be able to  help with the actual “history” part too. A 2021 study used neural networks to try  and identify if an ancient Hebrew scroll from the Dead Sea was written by a single person based on features of the handwriting.

And researchers found that it wasn’t  just a single person that wrote it; it was multiple scribes that were careful enough to have similar handwriting to each other! We are still in the early days for this kind  of work, but it’s beginning to look like the future of archeology could be a little  more “Tony Stark” than “Indiana Jones.” And that goes for the world of coffee,  too, thanks to today’s sponsor Cometeer. Cometeer brews barista-quality coffee  from the world’s best specialty roasters.

The coffee is ground, brewed,  and immediately flash-frozen to lock in the coffee bean’s  freshness and flavor. Then it gets delivered right to your door inside convenient recyclable aluminum capsules. And you can make the cometeer coffee  at home in just a couple of minutes, just add hot water, or cold if  you’d rather have iced coffee, and you’ll have a cup that tastes just  like a barista brewed it for you.

If you’re interested in trying them  out, they have a special offer going for fans of SciShow right now. You can get $20 off your first purchase plus free shipping when you click on our link. That’s ten free cups of coffee and over 30% off! [♪ OUTRO]