Regarding the discussion of ger/yurt districts in cities, it's also important not to underestimate the cultural significance of the nomadic lifestyle and yurt culture.
Changing climate (desertification) and economic conditions have meant that a lot of people have given up their nomadic lifestyle and moved to cities or their outskirts (mostly Ulaanbaatar). They often are reluctant to do so, it's a big step, and they often hope it is a temporary one.
They set up their yurts not only because of housing shortages, but many are also hesitant to move into apartments or other permanent structures as it's seen as the last step in giving up this nomadic lifestyle. Often they are setting up their yurts next to permanent structures, either because they are living in the 'yard' of relatives or to expand their residences and stay connected to their culture.
I've traveled across Mongolia on a motorcycle many years ago, and one thing I never expected is how absolutely everyone living in a permanent house also has a yurt in their backyard, regardless of how good the house is. This made no sense to me as an outsider (like, do you really need a second house?) so I asked a local about this, and was given a funny look. Yurts are just hardwired into the culture, it's a status symbol, it's where you invite a guest, it's what you use when living outside, it so many things at once.
One of my best Airbnb experiences was staying in a yurt in the backyard of a Mongolian woman…in Wisconsin. It was great. They also had a huge fire pit with tons of chairs around it, and I could tell they loved having tons of people over and just hanging out.
Only insofar as both building types are recognized externally as inextricably linked to the culture, right? Sauna is deeply rooted in Finnish culture but not quite to the level or multipurpose use of ger.
Agree - ger living is not necessarily a failure of public policy, it could just be a cultural decision. Even Genghis Khan lived in a ger. Of course, for some people, it's likely to be a matter of necessity, for others, a matter of choice, but it's not prima facie bad.
> When ineffective policy results in a large chunk of the populace generationally living in yurts on the outskirts of urban areas, it’s clear that there is failure.
The most annoying thing about modern life is people citing LLMs to try to win arguments about subjective questions. They are biased to agree with anything you ask them, and will do so unless it’s blatantly factually untrue.
Even if you ignore agreement, LLMs are trained on the content of the internet which is wildly biased toward the mean or lowest common denominator urban english speaking viewpoint depending on the subject.
The most annoying person in a chat group is the eternal LLM responder, that person that takes any question and feeds it to an LLM and replies back in the chat with it. We're now creating groups without these people to avoid the bullshit.
My pet theory is that LLM posters get more flak than they otherwise would because the equally terrible commenters who simply googled it, or worse, formed an opinion and googled up a cherry picked link to support it crowd feels threatened by them.
I'm not sure about Mongolia, but elsewhere I've seen wooden platforms for yurts under which water and sewage is ran, much as you would do with a crawlspace type house.
I've been in a Uzbekistan palace, I think it was Khiva. And it was, well, a palace, with courts and richly decorated rooms. But at certain fully enclosed by walls court there was a circular place where yurt have been standing. Khans were tracing their lineages back to Genghis Khan, and it was unbecoming for the khan to spend nights under the firm roof, even if it's in a middle of the city with long sedentary life style traditions. All the visiting relatives wouldn't approve.
Mongolia is also really struggling right now with a mass migration off the plains because of several very cold winters that have decimated their flocks. There just isn’t enough room for them to move into permanent buildings even if they wanted to.
It also sounds like they would already have one, and / or that it would be relatively easy to move if they want or need to. Don't they go back to their more rural homes for special events, for example?
Moving a ger can be a significant effort, especially a large one. Most urban dwellers find owning the trucks and vans that can hold these things pretty impractical if they're not moving regularly.
There's a fairly large domestic tourism industry catering to urban city-dwellers who want to go live in a nice ger for a couple weeks to feel connected to their history.
The gers are standardized. There is a big daily market in Ulaanbaatar where you can get all spare parts and complete gers. In 2017, the price for one ger was something like $1000.
For that money, you get a well-isolated easily movable tiny house in a country where you are allowed to settle everywhere (but if you have 2000 sheep with you, you should better discuss the usage of the pastureland with the locals) without paying rent (outside the city).
Choosing a ger for housing is not only about tradition and culture. It is quite rational in that situation.
Just to clarify for some people, in polite company, a turkish house-tent is what we call a yurt, whereas a house-tent in mongolia is not a yurt but a ger. In France these are called shabadoos, in canada we call house-tents plumbuses, and in the US these are called fleebs. I personally get offended when locals see my shabadoo and refer to it as a ger.
I’m a native speaker and the original phrasing was fine and sounds like completely correct idiomatic English to me.
Yes, the syntax is ambiguous, but ambiguously-parseable sentences happen all the time in all languages and we resolve the ambiguity using context clues, which in this case is easy to do.
The phrasing is correct and pretty normal, it’s just potentially ambiguous. English is like that sometimes. I’m not a grammarian, but I think “I counted all the yurts in Mongolia using machine learning” would normally be interpreted correctly by most people, with ‘using’ referring to the subject ‘I’. The way you’d write the other interpretation is “I counted all the yurts in Mongolia that use machine learning”. Your proposed alternative is also correct and less ambiguous.
"I counted all of the yurts in Mongolia using machine learning"
It's not wrong, but possibly ambiguous, and I'd bet an English teacher would prefer it was phrased differently. In speech, I wouldn't bat an eye at that arrangement. But, if I were to write the headline, it would have been...
"I used machine learning to count all the yurts in Mongolia."
As another commenter said, the phrasing isn't wrong, just ambiguous. I would add the word "by" to make it unambiguous: "I counted all of the yurts in Mongolia by using machine learning."
The quick fix without changing the sentence structure is to insert "by" before "using". Colloquial speech often omits it but it's implicit here. The other option is to switch to active voice.
It seems like a waste that you didn't use the 89,259 yurts that are already outlined in OpenStreetMap as input, though you would've probably had issues aligning the outlines with google maps imagery
> Finally, that's a much smaller number than I expected for a country of 3 million.
172.7k yurts. Assuming that these are family residences for the most part, if we take an average occupancy of 4 (which is probably too low - the fertility rate is still quite high there) gives ~691k people living in yurts - approximately 20% of the population of 3.5 million - sounds reasonable.
From my memory: 3 million people, 1.5 living in the capital.
Let's say 1 million are living outside cities.
4 people per yurt.
250,000 yurt.
Add some extra yurts because there will be people having more than one or people living in a house with a yurt in the garden or yurts used as warehouses, etc
300,000 which is almost the double of the count from the ML app.
This is a nice idea that often comes up in geo/ml projects. (Why not just use OSM for all your labels?)
To start, OSM doesn't use Google Maps imagery for annotation due to licensing concerns. As someone else mentioned, it's rarely clear whether researchers have the right to use Maps imagery let alone download/re-publish it. Part of the reason is that Google sub-licenses imagery from several different providers who are usually extremely protective of IP. So immediately you'd have image/label alignment issues.
Even if you had access to the image that someone used for labeling, it's non-trivial. They might not even have used an image! For example you might walk around and take a GPS reading next to every object and use the keypoints as object centers. Sometimes the annotation quality is low, for example if you want to try using building outlines or roads as segmentation targets for aerial imagery. Or things are simply misaligned. Also since yurts are inherently mobile, you might not even be able to use those labels because objects have moved and there's no guarantee they'll be present in Google Maps.
Finally you'd have issues of omission/commission, because you would have to assume that OSM is complete. That's very sensitive to how active the local community is. Some places are accurate down to the fire hydrant. Where I live, there are plenty of unmapped businesses that have been here for years. Though you could definitely use it to cross-check your own labels + predictions.
The standard for detecting objects on tiles is to discard border predictions and rely on overlap (sliding window) prediction + non max suppression (NMS) to handle duplicates. The overlap is usually something like 1x receptive field of your model, and your "discard" region is a bit larger than your max expected object size.
From experience, I agree with your points. One thing OSM data is ok for is land classification labels ("landuse", etc tags) as the accuracy is not as important at their scales and requires less effort to cleanup. Most of the work is aggregating disparate landuses into buckets that make sense for your model.
Wikipedia says 30% of 3.5 million are "nomadic or semi-nomadic", which would be 6 people to a yurt. I couldn't figure out what percentage of the country was done, but if he did 270,559/37,258,617 zoom 17 tiles then there could be another 100k in the other 99% of the data.
Living away from other people and not next to anything in particular is what I associate with nomads, the heuristic of searching a radius around landmarks doesn't make sense to me. I scrolled around a random remote desert area in Mongolia on Google Maps and found a yurt every couple of minutes.
PSA: Downloading Google Maps satellite imagery tiles is forbidden by the TOS. This is enforced, too, and I'm quite surprised the OP managed to download tiles for all of Mongolia without getting banned.
I don't think google owns lots of their tiles. If you view google maps on desktop you can see the copyright of an individual area, and it's often random aerospace companies.
Man, I had almost forgotten how much fun it was to read about ML projects of the variety I studied in university, before all the discourse shifted to plain English prompts.
I know of some government entities in Australia doing similar work, but the effectiveness/quality level of the author's work do make me despair for our government a bit. They're blowing years of Very Expensive Consultant spend and they can't even classify an entire parcel of land correctly, let alone count some little yurt-shaped blobs.
Object detectors output detection bounding boxes along with confidence scores. The higher the score, the more confident the model is that the associated bounding box is a correct detection.
When used in applications (like this one), the user typically establishes a confidence threshold and then every detection above that threshold is treated as a positive detection, the rest are discarded. The choice can be arbitrary or (sorta) principled.
With one extra step, you could get a lower bound with very high confidence. Just take a random sample (N=100) of the detected objects, and determine manually what percentage are really yurts. Then multiply that percentage by the number of detected objects.
All very cool, but wouldn't it be a lot easier to just detect circles within a certain range of size and color using some basic computer vision, like a circle Hough transform?
One potential issue with this is that gers actually leave circular patterns once they've been removed, where the soil is disturbed and the grass is light-starved or dead. It takes a couple years for those to fully disappear after the ger is removed. Gers themselves can be any color. White is traditional (from a time when white wool was more expensive), but not mandatory.
Ideally you’d verify against an in-person count of yurts over some control area. Otherwise this is just based on an assumption of what yurts look like on satellite.
It'd be a lot more accurate—not to say more honest—to say the author _estimated_ the number of all the yurts in Mongolia using machine learning. ML algorithms are stochastic; their outputs are whatever the algorithm deems the most probable of the options generated from the given inputs. They barely give a thought to all the ways their count could be wrong—no error analysis, no confidence intervals. There's a meaningless prediction score of 40%, and they blithely add "a hundred or so" to the count.
This is anti-information. People reading this uncritically will come away with completely wrong ideas about the number of yurts in Mongolia, about machine learning algorithms, about data science in general.
> People reading this uncritically will come away with completely wrong ideas about the number of yurts in Mongolia
Who is harmed by carrying around a mistaken number for this, especially if they notice the 40% confidence?
As to the rest, I read it as an application of tools for an interesting question, not a comprehensive or authoritative how-to. It’s scaled napkin math, and napkin math is very useful.
Assuming you selectively quoted me in good faith before asking "who is harmed", you should read the whole SEP entry on the ethics of manipulation, and then you should review the works it references.
- Whoever hires the author for their software engineering or data science expertise in part because of this blog post will pay for substandard work.
- By deceiving their audience as to the accuracy and precision of the demonstrated techniques, the author undermines the audience's ability to make good decisions about when to use or how to reason from the results of machine learning algorithms.
- The author disrespects their audience when they misrepresent themself, their work, and their results.
'yurt' is a Turkic word (don't confuse with Turkish), means "dwelling place", also commonly used for "home country" in languages like Uzbek. Yurt for the nomads was far more than just a place to sleep. Mongols I believe use a different word - 'ger'. Fun fact - Turkic languages were lingua franca of the Mongol Empire. Mongols perceived their language as "sacred" and wouldn't let non-mongols teach it freely. That could be one of the reasons why we have Turkic languages spread across the vast territory instead of Mongolic.
Just so you know, Turkic languages span an enormous landmass from Turkey in the west, across the Caucasus, Central Asia, Russia, China, parts of Iran, Afghanistan and Mongolia. This represents one of the largest continuous language family distributions on Earth - spanning roughly 13,000+ kilometers east-west across Eurasia.
Regarding the discussion of ger/yurt districts in cities, it's also important not to underestimate the cultural significance of the nomadic lifestyle and yurt culture.
Changing climate (desertification) and economic conditions have meant that a lot of people have given up their nomadic lifestyle and moved to cities or their outskirts (mostly Ulaanbaatar). They often are reluctant to do so, it's a big step, and they often hope it is a temporary one.
They set up their yurts not only because of housing shortages, but many are also hesitant to move into apartments or other permanent structures as it's seen as the last step in giving up this nomadic lifestyle. Often they are setting up their yurts next to permanent structures, either because they are living in the 'yard' of relatives or to expand their residences and stay connected to their culture.
You can see examples of this in the first images.
I've traveled across Mongolia on a motorcycle many years ago, and one thing I never expected is how absolutely everyone living in a permanent house also has a yurt in their backyard, regardless of how good the house is. This made no sense to me as an outsider (like, do you really need a second house?) so I asked a local about this, and was given a funny look. Yurts are just hardwired into the culture, it's a status symbol, it's where you invite a guest, it's what you use when living outside, it so many things at once.
One of my best Airbnb experiences was staying in a yurt in the backyard of a Mongolian woman…in Wisconsin. It was great. They also had a huge fire pit with tons of chairs around it, and I could tell they loved having tons of people over and just hanging out.
So it's basically Mongolia's answer to the Finnish sauna
Only insofar as both building types are recognized externally as inextricably linked to the culture, right? Sauna is deeply rooted in Finnish culture but not quite to the level or multipurpose use of ger.
sounds like the concept of a pool/guest house to me
Agree - ger living is not necessarily a failure of public policy, it could just be a cultural decision. Even Genghis Khan lived in a ger. Of course, for some people, it's likely to be a matter of necessity, for others, a matter of choice, but it's not prima facie bad.
> When ineffective policy results in a large chunk of the populace generationally living in yurts on the outskirts of urban areas, it’s clear that there is failure.
That's not at all clear.
> That's not at all clear.
LLMs agree with OP. It's a failure, with important culture.
Steelmanning it, it's better than a corrugated metal shanty town. Although they would die in the cold.
The rich in the gers burn coal, the poor plastic. There is no water or sewerage.
It's one of the most polluted capitals in the world - https://www.unicef.org/mongolia/environment-air-pollution#:~...
Ulaanbaatar - https://geohack.toolforge.org/geohack.php?params=47_56_7_N_1...
I bet I could talk an LLM into supporting ger life. Probably in 40 words’ effort or less.
[dead]
What’s the point of citing unspecified “LLMs”? Do you expect this to be persuasive? And why more than one?
[dead]
The most annoying thing about modern life is people citing LLMs to try to win arguments about subjective questions. They are biased to agree with anything you ask them, and will do so unless it’s blatantly factually untrue.
Even if you ignore agreement, LLMs are trained on the content of the internet which is wildly biased toward the mean or lowest common denominator urban english speaking viewpoint depending on the subject.
The most annoying person in a chat group is the eternal LLM responder, that person that takes any question and feeds it to an LLM and replies back in the chat with it. We're now creating groups without these people to avoid the bullshit.
My pet theory is that LLM posters get more flak than they otherwise would because the equally terrible commenters who simply googled it, or worse, formed an opinion and googled up a cherry picked link to support it crowd feels threatened by them.
(both groups are trash, IMO)
Please do not call people "trash" here.
I'm not sure about Mongolia, but elsewhere I've seen wooden platforms for yurts under which water and sewage is ran, much as you would do with a crawlspace type house.
[dead]
I've been in a Uzbekistan palace, I think it was Khiva. And it was, well, a palace, with courts and richly decorated rooms. But at certain fully enclosed by walls court there was a circular place where yurt have been standing. Khans were tracing their lineages back to Genghis Khan, and it was unbecoming for the khan to spend nights under the firm roof, even if it's in a middle of the city with long sedentary life style traditions. All the visiting relatives wouldn't approve.
https://en.wikipedia.org/wiki/Toshhovli_Palace
https://en.wikipedia.org/wiki/Toshhovli_Palace#/media/File:K... that circular spot.
Mongolia is also really struggling right now with a mass migration off the plains because of several very cold winters that have decimated their flocks. There just isn’t enough room for them to move into permanent buildings even if they wanted to.
It also sounds like they would already have one, and / or that it would be relatively easy to move if they want or need to. Don't they go back to their more rural homes for special events, for example?
Moving a ger can be a significant effort, especially a large one. Most urban dwellers find owning the trucks and vans that can hold these things pretty impractical if they're not moving regularly.
There's a fairly large domestic tourism industry catering to urban city-dwellers who want to go live in a nice ger for a couple weeks to feel connected to their history.
The gers are standardized. There is a big daily market in Ulaanbaatar where you can get all spare parts and complete gers. In 2017, the price for one ger was something like $1000.
For that money, you get a well-isolated easily movable tiny house in a country where you are allowed to settle everywhere (but if you have 2000 sheep with you, you should better discuss the usage of the pastureland with the locals) without paying rent (outside the city).
Choosing a ger for housing is not only about tradition and culture. It is quite rational in that situation.
Just to clarify for some people, in polite company, a turkish house-tent is what we call a yurt, whereas a house-tent in mongolia is not a yurt but a ger. In France these are called shabadoos, in canada we call house-tents plumbuses, and in the US these are called fleebs. I personally get offended when locals see my shabadoo and refer to it as a ger.
Do they build some kind of foundation for them?
Depends. Permanent ones, and tourist gers, yes. Actual nomadic gers are just placed on the grass with rugs.
There are zero yurts in Mongolia using machine learning.
I'd bet quite highly that it's non-zero.
The yurts themselves? Seems unlikely. Someone living in a yurt? Absolutely.
I assumed a yurt was a type of person/job, so I initially read the title the same way
thank you for this
*edit (I mean this sincerely, it made me laugh and I did not see it at first)
I chuckled a little, but as a non-native speaker: what would be the correct phrasing? "Using machine learning, I counted all the yurts in Mongolia?"
I’m a native speaker and the original phrasing was fine and sounds like completely correct idiomatic English to me.
Yes, the syntax is ambiguous, but ambiguously-parseable sentences happen all the time in all languages and we resolve the ambiguity using context clues, which in this case is easy to do.
Ambiguity is the soul of wit, in this case
Soul of twit, more like.
The phrasing is correct and pretty normal, it’s just potentially ambiguous. English is like that sometimes. I’m not a grammarian, but I think “I counted all the yurts in Mongolia using machine learning” would normally be interpreted correctly by most people, with ‘using’ referring to the subject ‘I’. The way you’d write the other interpretation is “I counted all the yurts in Mongolia that use machine learning”. Your proposed alternative is also correct and less ambiguous.
"I counted all of the yurts in Mongolia using machine learning"
It's not wrong, but possibly ambiguous, and I'd bet an English teacher would prefer it was phrased differently. In speech, I wouldn't bat an eye at that arrangement. But, if I were to write the headline, it would have been...
"I used machine learning to count all the yurts in Mongolia."
good one.
or even just add a comma at the right place:
"I counted all of the yurts in Mongolia, using machine learning"
As another commenter said, the phrasing isn't wrong, just ambiguous. I would add the word "by" to make it unambiguous: "I counted all of the yurts in Mongolia by using machine learning."
Just add a comma. "I counted all of the yurts in Mongolia, using machine learning"
Unfortunately, that isn't a correct use of a comma
In this case, "using machine learning" is an independent clause, so the comma is required. Without the comma, it modifies "yurts in Mongolia".
...I don't think that's right? Not an expert.
Better yet, "I used machine learning to count all the yurts in Mongolia"
This definitely doesn’t flow better in English.
https://www.amazon.com/Eats-Shoots-Leaves-Tolerance-Punctuat...
The quick fix without changing the sentence structure is to insert "by" before "using". Colloquial speech often omits it but it's implicit here. The other option is to switch to active voice.
I have no idea why 15 people decided it was important to repeat that it’s fine rather than answering your question.
Yes, when the phrase is ambiguous, it’s usually more coherent to simply change the sentence order as you’ve done here.
the phrasing is fine, is just that that sort of construction is a common source of humour in english. one famous example is groucho marx's
"One morning I shot an elephant in my pajamas. How he got into my pajamas I'll never know."
[dead]
It seems like a waste that you didn't use the 89,259 yurts that are already outlined in OpenStreetMap as input, though you would've probably had issues aligning the outlines with google maps imagery
https://taginfo.geofabrik.de/asia:mongolia/tags/building=ger
I'm also guessing your model doesn't handle yurts that are on the border of a tile.
Finally, that's a much smaller number than I expected for a country of 3 million.
> Finally, that's a much smaller number than I expected for a country of 3 million.
172.7k yurts. Assuming that these are family residences for the most part, if we take an average occupancy of 4 (which is probably too low - the fertility rate is still quite high there) gives ~691k people living in yurts - approximately 20% of the population of 3.5 million - sounds reasonable.
About 45% in 2010 apparently https://unstats.un.org/unsd/demographic-social/census/docume...
My quick estimate before clicking the link was:
From my memory: 3 million people, 1.5 living in the capital.
Let's say 1 million are living outside cities.
4 people per yurt.
250,000 yurt.
Add some extra yurts because there will be people having more than one or people living in a house with a yurt in the garden or yurts used as warehouses, etc
300,000 which is almost the double of the count from the ML app.
This is a nice idea that often comes up in geo/ml projects. (Why not just use OSM for all your labels?)
To start, OSM doesn't use Google Maps imagery for annotation due to licensing concerns. As someone else mentioned, it's rarely clear whether researchers have the right to use Maps imagery let alone download/re-publish it. Part of the reason is that Google sub-licenses imagery from several different providers who are usually extremely protective of IP. So immediately you'd have image/label alignment issues.
Even if you had access to the image that someone used for labeling, it's non-trivial. They might not even have used an image! For example you might walk around and take a GPS reading next to every object and use the keypoints as object centers. Sometimes the annotation quality is low, for example if you want to try using building outlines or roads as segmentation targets for aerial imagery. Or things are simply misaligned. Also since yurts are inherently mobile, you might not even be able to use those labels because objects have moved and there's no guarantee they'll be present in Google Maps.
Finally you'd have issues of omission/commission, because you would have to assume that OSM is complete. That's very sensitive to how active the local community is. Some places are accurate down to the fire hydrant. Where I live, there are plenty of unmapped businesses that have been here for years. Though you could definitely use it to cross-check your own labels + predictions.
The standard for detecting objects on tiles is to discard border predictions and rely on overlap (sliding window) prediction + non max suppression (NMS) to handle duplicates. The overlap is usually something like 1x receptive field of your model, and your "discard" region is a bit larger than your max expected object size.
From experience, I agree with your points. One thing OSM data is ok for is land classification labels ("landuse", etc tags) as the accuracy is not as important at their scales and requires less effort to cleanup. Most of the work is aggregating disparate landuses into buckets that make sense for your model.
> Finally, that's a much smaller number than I expected for a country of 3 million.
172k of them? That still seems like quite a lot of yurts; certainly more yurts per capita than anyone else has.
Wikipedia says 30% of 3.5 million are "nomadic or semi-nomadic", which would be 6 people to a yurt. I couldn't figure out what percentage of the country was done, but if he did 270,559/37,258,617 zoom 17 tiles then there could be another 100k in the other 99% of the data.
Living away from other people and not next to anything in particular is what I associate with nomads, the heuristic of searching a radius around landmarks doesn't make sense to me. I scrolled around a random remote desert area in Mongolia on Google Maps and found a yurt every couple of minutes.
I'm confused why you wouldnt just do some random sampling to get some statistical bounds. At least then you'll know if you are close.
PSA: Downloading Google Maps satellite imagery tiles is forbidden by the TOS. This is enforced, too, and I'm quite surprised the OP managed to download tiles for all of Mongolia without getting banned.
I don't understand the reasoning behind that besides market exclusivity.
I don't think google owns lots of their tiles. If you view google maps on desktop you can see the copyright of an individual area, and it's often random aerospace companies.
Ah, makes sense! Thanks
Just create another account.
Right? Who gives a fuck.
Maxar et. al
Intrigued by this. What was the rate of false positives? For example are there storage tanks, silos, above ground pools mistaken for yurts?
Man, I had almost forgotten how much fun it was to read about ML projects of the variety I studied in university, before all the discourse shifted to plain English prompts.
I know of some government entities in Australia doing similar work, but the effectiveness/quality level of the author's work do make me despair for our government a bit. They're blowing years of Very Expensive Consultant spend and they can't even classify an entire parcel of land correctly, let alone count some little yurt-shaped blobs.
I was just about to mention that there's a local developer/improv artist with similar experiences and then I saw your username
"In total I found 172,689 yurts with a prediction score of greater than 40%."
How should one interpet the "prediction score"?
Object detectors output detection bounding boxes along with confidence scores. The higher the score, the more confident the model is that the associated bounding box is a correct detection.
When used in applications (like this one), the user typically establishes a confidence threshold and then every detection above that threshold is treated as a positive detection, the rest are discarded. The choice can be arbitrary or (sorta) principled.
Ok, then "prediction score" is the confidence score? And the confidence threshold for an artefact being a yurt is 40%?
I thought the headline was discussing all of the yurts in Mongolia that are using machine learning.
They use a semi-commercial solution (free for educational use).
I'm curious what the topology/architecture of the DL model is like. And are there better ways to approach this problem?
Keeping an eye on the steppe nomads is always a good idea.
Using machine learning, he counted all of the yurts.
Counting all of the yurts that happen to be using machine learning is a way more difficult problem.
No validation versus some kind of ground truth . His training data set is very small and geographically limited. His model is likely pretty inaccurate
How much approximately does it cost to rent these amounts of gpu + server?
Nice write up, also great to see Docker Swarm being used.
Nice, thanks for sharing! What would be the best way (and data source) to observe the number of yurts over time?
You estimated. This is not at all a "count."
With one extra step, you could get a lower bound with very high confidence. Just take a random sample (N=100) of the detected objects, and determine manually what percentage are really yurts. Then multiply that percentage by the number of detected objects.
Maybe I missed it but did the article give an answer for the number of yurts?
It's in the article, second to last passage: "In total I found 172,689 yurts with a prediction score of greater than 40%."
All very cool, but wouldn't it be a lot easier to just detect circles within a certain range of size and color using some basic computer vision, like a circle Hough transform?
One potential issue with this is that gers actually leave circular patterns once they've been removed, where the soil is disturbed and the grass is light-starved or dead. It takes a couple years for those to fully disappear after the ger is removed. Gers themselves can be any color. White is traditional (from a time when white wool was more expensive), but not mandatory.
Where was this when Morrowind came out? :)
Nice! Now how will you validate the result?
Ideally you’d verify against an in-person count of yurts over some control area. Otherwise this is just based on an assumption of what yurts look like on satellite.
[dead]
What would this have cost?
It'd be a lot more accurate—not to say more honest—to say the author _estimated_ the number of all the yurts in Mongolia using machine learning. ML algorithms are stochastic; their outputs are whatever the algorithm deems the most probable of the options generated from the given inputs. They barely give a thought to all the ways their count could be wrong—no error analysis, no confidence intervals. There's a meaningless prediction score of 40%, and they blithely add "a hundred or so" to the count.
This is anti-information. People reading this uncritically will come away with completely wrong ideas about the number of yurts in Mongolia, about machine learning algorithms, about data science in general.
> People reading this uncritically will come away with completely wrong ideas about the number of yurts in Mongolia
Who is harmed by carrying around a mistaken number for this, especially if they notice the 40% confidence?
As to the rest, I read it as an application of tools for an interesting question, not a comprehensive or authoritative how-to. It’s scaled napkin math, and napkin math is very useful.
Assuming you selectively quoted me in good faith before asking "who is harmed", you should read the whole SEP entry on the ethics of manipulation, and then you should review the works it references.
https://plato.stanford.edu/entries/ethics-manipulation/
But to answer you directly:
- Whoever hires the author for their software engineering or data science expertise in part because of this blog post will pay for substandard work.
- By deceiving their audience as to the accuracy and precision of the demonstrated techniques, the author undermines the audience's ability to make good decisions about when to use or how to reason from the results of machine learning algorithms.
- The author disrespects their audience when they misrepresent themself, their work, and their results.
this is 1 step above yak shaving i suppose
So hop on the bandwagon before you get left behind by Mongolians in yurts.
Cool, how can this be used for taxation purposes?
[dead]
[flagged]
[flagged]
Yurt is a lot of fun to say. Great word.
'yurt' is a Turkic word (don't confuse with Turkish), means "dwelling place", also commonly used for "home country" in languages like Uzbek. Yurt for the nomads was far more than just a place to sleep. Mongols I believe use a different word - 'ger'. Fun fact - Turkic languages were lingua franca of the Mongol Empire. Mongols perceived their language as "sacred" and wouldn't let non-mongols teach it freely. That could be one of the reasons why we have Turkic languages spread across the vast territory instead of Mongolic.
Just so you know, Turkic languages span an enormous landmass from Turkey in the west, across the Caucasus, Central Asia, Russia, China, parts of Iran, Afghanistan and Mongolia. This represents one of the largest continuous language family distributions on Earth - spanning roughly 13,000+ kilometers east-west across Eurasia.