How Deepmind AI Can Help Us Learn from Our Ancient History?
DeepMind, a London-based artificial intelligence research group owned by Google's parent company Alphabet.
Updated
On this page
Updated November 2026. This was first published on 22 March 2022. The idea is the same. Any number below is what the original piece said then. It is not a new count.
Key takeaways
- DeepMind, a London-based artificial intelligence research group owned by Google’s parent company Alphabet, has built an artificial intelligence program that beats similar software in many linguistic tasks.
- Artificial intelligence and machine learning techniques like Deepmind speed up such studies and bring attention to previously neglected data.
- For example, this artificial intelligence technology, which was created in partnership with historians and DeepMind, called Ithaca, can assist in recreating ancient Greek writings with 72% accuracy.
DeepMind, a London-based artificial intelligence research group owned by Google’s parent company Alphabet, has built an artificial intelligence program that beats similar software in many linguistic tasks.
Artificial intelligence and machine learning techniques like Deepmind speed up such studies and bring attention to previously neglected data. For example, this artificial intelligence technology, which was created in partnership with historians and DeepMind, called Ithaca, can assist in recreating ancient Greek writings with 72% accuracy.
More than 63,000 transcribed ancient Greek inscriptions were fed into Ithaca, allowing it to identify patterns in the arrangement of letters and words, correlations between words and phrases, and the age and origin of a document.
The AI system can also pinpoint with greater than 70% accuracy where the writings were written in the ancient Mediterranean and date them to within a few decades of their agreed-upon origination date. This is an advancement above the AI’s previous iteration, which could only reconstruct ancient writings.
How Has This Helped the Historians Solve Problems? #
Epigraphers use large libraries of material to create textual and contextual similarities while recreating damaged inscriptions. Parallel mnemonic repertoires and, more recently, digital corpora for string matching searches are commonplace in these repositories.
It’s impossible to determine the underlying probability distribution of potential restorations if the search query differs from what is really being searched for.
Researchers are increasingly dealing with digitized data rather than sources that have been manually curated into archives. Historians’ views and findings will be corrupted if they do not modify their techniques of searching and analyzing information.
Physical documents, for example, maybe digitized using an AI technology known as optical character recognition. But, letters and words are sometimes mistaken or completely absent even when the accuracy is excellent. Historical data may lead historians astray at a more fundamental level if algorithms are employed to determine what should be archived and what should be removed.
The task of restoring or reproducing ancient relics is a difficult one. Models of computer vision previous work ? For example, the deterioration of images or papers has not been accounted for.
We can overcome the limits of current epigraphic techniques by merging new machine learning research with a new AI (named Ithaca). Complex statistical patterns can be discovered and exploited by deep neural networks based on biological neural networks.
Is It Efficient? #
According to testing, Ithaca can restore broken text with 62 percent accuracy on its own, compared to 25 percent accuracy for human historians. But, combining man and machine improves total accuracy to 72 percent, demonstrating “the potential for human-machine collaboration” in the field; Ithaca can date inscriptions within 30 years and attribute them to their original place with a 71 percent accuracy.
Ithaca has already shown its worth to historians in a test case involving a series of Athenian regulations that have been the subject of a dating debate. According to historians, the decrees were initially dated to 446 BC. This evaluation was based on particular letterforms employed by the Athenian bureaucracy at the time (known as the Attic three-bar sigma). The Athenians began using an Ionic four-bar sigma for their decrees around 446 BC.
The researchers emphasized the model’s adaptability and claimed it could be applied to any ancient language from Latin to Mayan to Cuneiform. It may also be able to train the system on Greek literary works written on papyrus pieces. might provide insight on poets like Sappho’s writings.
How Does the AI Tool Work on These Ancient Texts? #
The AI model learned how to reverse the erosion process and estimate what original fragments looked like before attempting to recreate artifacts. The researchers then devised a method for determining whether pieces fit together in the model.
Historians may better assess the AI’s predictions by giving weightings to input text elements used to forecast missing characters. In other situations, the method aids in improving neural networks with transparent decision-making processes.
Transform22 is a neural network architecture that uses an attention mechanism to balance the impact of various input items (such as letters and words) on the model’s decision-making process. is how Ithaca’s torso is constructed.
For this novel attention method, each component of the incoming text is concatenated with its sequential positioning information.
Stacks of transformer blocks make up the structure of Ithaca’s body, and each one outputs a series of processed representations that is exactly as long as the number of letters entered into it. When a block is finished, its output is used as the input for the next. The torso’s final output is sent to three task heads, each responsible for the restoration, geographical attribution, and chronological attribution.
AI Like Ithaca is Essential for Our World #
What is kept out of history has real-world implications in healthcare, economics, and international affairs. Moreover, people and groups with less power and privilege are disproportionately affected.
Machine learning approaches like Ithaca are best suited to uncovering gaps in historical documents, but historians must exercise caution when using data science tools.
While promising thus far, it’s vital to remember that AI models like Ithaca can’t operate without the help of humans. Ithaca and historians collaborated on the most accurate reconstructions. When historians employed Ithaca’s top 20 most likely reconstructions for a particular text to guide their work, they could recover the text with more accuracy than Ithaca alone. Ithaca’s design should allow it to be translated into any ancient language and published in any medium.
More testing with historians is required, and users will require training and technical help to use the program. We have a long way to go in using AI’s like Ithaca.
Questions and answers #
How Has This Helped the Historians Solve Problems? #
Epigraphers use large libraries of material to create textual and contextual similarities while recreating damaged inscriptions. Parallel mnemonic repertoires and, more recently, digital corpora for string matching searches are commonplace in these repositories.
Is It Efficient? #
According to testing, Ithaca can restore broken text with 62 percent accuracy on its own, compared to 25 percent accuracy for human historians.
How Does the AI Tool Work on These Ancient Texts? #
The AI model learned how to reverse the erosion process and estimate what original fragments looked like before attempting to recreate artifacts. The researchers then devised a method for determining whether pieces fit together in the model.