So far, the project has produced amazing results. A pattern found in the data allowed researchers to see how scientific knowledge was gathering as Europe was torn along religious lines after the Protestant Reformation. Scientific writings being published in places such as Protestant Wittenberg, a center of scholarly innovation due to the work of the Reformers, were being imitated in centers such as Paris and Venice before spreading across the continent. The Protestant Reformation is not a well-studied subject, Valeriani says, but the machine-centric perspective has allowed researchers to see something new: “This was never clear before.” Models applied to tables and figures are beginning to return similar patterns.
Computers are often only modern iterations of things with longer histories—think iPhones and Teslas rather than Switchboards and Model Ts.
These tools offer more opportunities than simply keeping track of 10,000 charts, Valeriani says. Instead, it allows researchers to draw inferences about the evolution of knowledge from patterns in clusters of documents, even if they examine only a handful of documents. “Looking at two tables, I can make a big conclusion about 200 years,” he said.
Deep neural networks are playing a role in analyzing even the past. Deciphering inscriptions (known as epigraphy) and restoring damaged examples are laborious tasks, especially when the inscriptions have been moved or contextual clues are missing. Special historians must make educated guesses. To help, DeepMind research scientist Yanis Assail and Tia Somerschild, a postdoctoral fellow at Venice’s Ca’ Foscari University, created a neural network called Ithaca, which can reconstruct parts of the missing texts and pinpoint the dates and locations of the texts. The deep learning approach, which involves training on a dataset of more than 78,000 articles, is the first to learn from large amounts of data and jointly solve regression and formation, the researchers said.
So far, Assael and Somerscheld say, this approach is shedding light on decrees written during the critical period in ancient Athens, which some historians have long argued for between 446 and 445 BCE. As an experiment, researchers trained the model on a data set that did not contain the text in question, and then asked it to analyze the text of the announcements. This set a different day. “The average date of the Ithaca edicts is predicted to be 421 BC, which aligns with recent dating findings and shows that machine learning can contribute to the debate over one of the most important periods in Greek history,” they said in an email.
Beth Hokel
Time machines
Other projects propose using machine learning to draw broader inferences about the past. This was the motivation behind the Venice Time Machine, one of several local “time machines” in Europe now used to reconstruct local history from digital archives. The archives of the Venetian Empire cover 1,000 years of history spread over 80 kilometers of shelves; The aim of the researchers was to digitize these records, most of which have not been examined by modern historians. Using deep learning networks to extract information and search for names found in the same document in other documents, they reconstruct the ties that once bound the Venetians.
Frederic Kaplan, president of the Time Machine Company, said that the project had digitized the administrative documents of the city over the centuries, capturing the image of the city and building and identifying the families that lived in different areas. Points in time. “To get to this level of flexibility, hundreds of thousands of these documents have to be digitized,” says Kaplan. “This has never been done before.”
Still, when it comes to the ultimate hope of the project—no less than a digital simulation of medieval Venice at the neighborhood level, with networks reconstructed by artificial intelligence—historians like Johannes Prezer-Kappeler, a professor at the Austrian Academy of Sciences who led the study. According to the Byzantine bishops, the project failed to deliver because the model failed to understand which connections made sense.
