Trajectory #13: Machine Learning and Archaeology
Silicon Valley Meets Mesopotamia
The High-Level This Week: Machine learning is transforming archaeology. Everything from site discovery and identification, artifact classification and document translation will be disrupted by new machine learning-driven approaches to archaeology. The next few decades will see a significant acceleration in the pace of archaeological discovery.
To learn more, read on.
If you haven’t already, subscribe here. If you liked this week’s edition and think someone else you know would find it interesting please take a couple seconds to share it via email or on social media.
The Perverse Panic over Plastic: Is plastic use really that bad?
Food And Agtech Startups Raised $20 Billion In 2019, Led By Meat Alternatives And Indoor Agriculture: Trajectory #4 covered discussed a portion of the increasing innovation happening in food today.
Hawley seeks overhaul of the Federal Trade Commission: An interesting proposal to strip the FTC of antitrust powers.
Sickeningly Sweet: Evidence is growing that sugar and cancer have a special relationship.
United Airlines buys flight-training academy to speed up hiring of 10,000 pilots: United Airlines is doing its best to alleviate the impending pilot shortage.
Key Issue: Archaeological Machine Learning
Archeology is a relatively obscure but deeply impactful discipline. Archaeology is critical to our understanding of human history. Moreover, the discoveries unearthed through archaeological work impact and help shape other intellectual disciplines such as politics, geography, demographics and more. In many ways archaeology is a foundational discipline that creates the base of data and historical context that many different fields rely on.
Archaeology has three core functions: surveying, excavation and analysis. Archaeologists work diligently to identify archeological sites, excavate them to find artifacts and structures and then analyze them. Archaeology is a resource-intensive, slow and highly manual field. Archaeological teams can number in the hundreds and archeological digs can last for years. For example over 1000 archaeologists will be working across 60 sites in the United Kingdom’s biggest archaeological excavation ever. The speed at which we can identify, excavate and analyze has long limited the rate at which archaeology – and in turn our understanding of human history – progresses.
However, a range of new archaeological techniques involving the application of machine learning to the archaeological process are being developed that promise to uncuff archaeology’s advancement from many of its historical limitations. Archaeological Machine Learning promises to be a critical and historic inflection point for the field of archaeology. In particular machine learning can help unlock critical bottlenecks in the surveying and analysis functions of archaeology. Excavation is likely to remain highly manual because of the delicacy required to handle historical artifacts.
Archaeological Site Prediction
One of the most difficult, slow and time-intensive tasks in archaeology is archaeological site prediction. In fact, archaeologists work diligently for years trying to identify new archaeological sites that are worthy of the time and cost to excavate. Traditionally archaeologists would manually pore over data to try identify sites worth of discovery. However, satellite imaging and LIDAR combined with machine learning promise to automate large parts of this discovery process.
LIDAR is a surveying method that involves using laser light to illuminate a target and measure the reflect light with a sensor. Differences in laser return times and wavelengths can be used to create 3D representations of targets. Lidar has long been used by archaeologists. However, machine learning is starting to allow archaeologists to automate the analysis of LIDAR and satellite data to automatically predict and identify archeological sites.
For example, researchers recently built a machine learning model that analyzed satellite imagery and correctly predicted 94% of the known archeological sites in Syria. Incredibly, the model also identified another 14,000 potential sites – more than 18 times the number of sites that have been identified by archaeologists in the region. Similarly and even more amazingly, using satellite imagery and machine learning Japanese researchers were able to identify 143 previously undiscovered Nazca lines!
Satellite image of some of the Nazca lines discovered by researchers.
Automated Artifact Classification
Another extremely manual task in archaeology is artifact classification. Every successful archaeological dig results in numerous archaeological artifacts which researchers then have to classify and analyze. Typically this process is done manually via visual inspection. For example, for classification of historical glass artifacts, an expert will search through a reference collection to try classify new glass pieces that are discovered.
Another example involves the classification of coins that are discovered. Once again, experts use a reference collection to try classify new coins. These classification problems can be automated in a significant way via machine learning. A recent paper walks through how researchers built a machine learning model for automated coin classification based off of a training dataset of 692 modern European coins. When researchers tested the model, they found that the model classified 78% of the coins in the test dataset correctly. They found the miscalculations were usually due to some of the coins being dirty or simply of unknown origin not flaws in the model.
Automated Artifact Translation
Archaeologists have also started to use machine learning to understand and recreate ancient Greek texts from broken stone tablets. DeepMind – Google’s famous deep learning unit – built a system called Pythia trained to recreate these texts. As a test PhD students and Pythia were both given a set of texts with artificially removed portions and asked to fill in the gaps. The students completed the text with a poor 43% accuracy. In contrast, Pythia correctly filled in the gaps a more impressive 70% of the time.
Machine learning can also be applied to translate ancient texts. For example, archaeologists have found thousands of Mesopotamian tablets over the years. However, most of these tablets remain undeciphered because of the limited number of experts fluent in these Mesopotamian languages and the time it takes to translate a document. Over 90% of these documents remain untranslated and they represent a treasure trove of historical information waiting to be explored. Machine learning can help by dramatically speeding up this translation process once researchers build correctly translated models.
Automated Archaeological Puzzle-Solving
Finally, machine learning can help researchers significantly accelerate the archaeological puzzle-solving process. Initially when most artifacts are discovered they are often broken into multiple pieces. Archaeologists spend countless hours figuring out how to reassemble these pieces. This puzzle-solving process can be incredibly taxing. Researchers at the University of Haifa have trialed a machine learning model that can predict how to reassemble fragmented artifacts for archaeologists. Initial testing has shown the model they built to perform extremely well on a few tests artifacts.
Archaeological Machine Learning
Overall, machine learning promises to drive a dramatic increase in the pace of archaeological research globally by unlocking a few critical bottlenecks. By automating and speeding up tasks such as archaeological site prediction, artifact classification, translation and puzzle-solving, machine learning will open the floodgates for archaeological research. Our understanding of human history and its complexities will advance rapidly in the decades to come as machine learning helps us discover previously undiscovered sites, classify and translate the artifacts we finds and enable archaeologists to focus on the most high-value and difficult archaeological tasks.