Hacker News Reader

Comment by AlotOfReading | Hacker News Reader

They have. Fancy statistical models are getting pretty common. It still requires funding and ground truthing by archaeologists, which are the real bottlenecks.

View on HN

I mean like telling the computer, “Show me more places that look like this“. Then, getting a bunch of results, saying yes or no to them as potential candidates, and watching the model get better and better showing you examples which you will confirm to be suitable candidates.

Yes, that's what I said. The term of art is "archaeological predictive modeling", but basically it's just finding statistical correlations between sites and spatial variables like soil type, water availability, etc. It's been around since the 80s, but a more recent example with ML is [1]. The people who find this most useful are heritage organization / BLM types that want to evaluate the "archaeological risk" of developing land without the need to do expensive surveys.

They're not nearly as useful in research archaeology because research is less focused on finding sites similar to what we already know about and more focused on getting an accurate picture of everything, including the sites not captured by traditional survey methodologies. Predictive modeling has a tendency reproduce your input biases, requiring careful management. A good paper that deals with this sort of problem and illustrates the scale of the issue is [2].

[1] https://doi.org/10.1371/journal.pone.0239424

[2] https://doi.org/10.7183/2326-3768.2.3.222

Looks like they are picking a set of variables. That’s a bit like looking at your past girlfriends, picking some of their features (hair color, hobbies), and then trying to predict from that which woman you’re going to fall in love with next. Of course, life will often surprise you in very unexpected ways.

I really just meant looking at satellite images for submerged cities just a few yards offshore, just based on visual similarity.

That happens too. However, archaeology tends to run on budgets that can be generously described as "shoestring". Most of the projects I worked on (some years-long), had total budgets less than half my current tech-company compensation. Since I like to work with older stuff, you can look at the sea level chart [1] and see obvious issue. Submersibles just aren't feasible for most research.

[1] https://upload.wikimedia.org/wikipedia/commons/1/1d/Post-Gla...

The problem is that you have to go examine (dig/dive/whatever) the potential sites to get real feedback -- otherwise, you're just training the model based on your own prejudices about what good sites look like.

True. And by going there, you might improve your understanding of what good candidate sites look like. Then again, having a longer list of candidates to choose from for such missions might still be helpful.

Reply on news.ycombinator.com