科技食谱

古希腊铭文,AI在几秒钟内解码

DeepMind, a Google-affiliated artificial intelligence development company, is increasing the utilization of AI algorithms such as disease prediction and energy efficiency. This Deep Mind announced on October 15 that it had developed an algorithm that predicts missing explanations in ancient inscriptions. Human experts say that a task that takes two hours can be done in seconds with AI.

To understand ancient society and culture, it is necessary to decipher the sentences on scrolls and ruins. However, these documents are usually deliberately destroyed or scattered over time as they weathered, or remain incomplete. In the study of inscriptions, experts are attempting to decipher these incomplete documents, but the work is complicated and time consuming.

The DeepMind research team, led by Yannis Assael, trained a neural network using Pythia to learn language patterns such as context, grammar, and sentence layout from 35,000 artifacts containing 3 million words. It is said that if Patia, who has learned the language pattern, is given an inscription, they can present about 20 candidates to fill in the empty seats. This technology allows experts to choose the right one from the alternatives proposed by Patia without having to think to fill the gap from scratch.

In fact, as a result of conducting an inscription supplement test to confirm the ability of Fatia, it was found that Fatia made 30% less mistakes than humans. In addition, it takes 2 hours for humans to fill 50, but Fatia completes it in seconds.

There are steps that still require human hands as well as neural networks to fully decipher the inscription. However, an expert at Oxford University who participated in the study said that the inscriptions speak of all aspects of the ancient world, including religion, society, and economy. It is easy for human experts to overlook details when decoding long sentences, but it can be seen that the new algorithm has a high success rate in decoding such sentences. Related information can be found here .