5 ms·

Pigeons taught to diagnose breast cancer on X-rays

35 points by randomname2 11y ago

brudgers 11y ago
Direct link to paper: http://www.plosone.org/article/fetchObject.action?uri=info:doi/10.1371/journal.pone.0141357&representation=PDF http://www.plosone.org/article/fetchObject.action?uri=info:d...
yareally 11y ago
Back in the 60s, there was a study done with pigeons being able to inspect pills on an assembly line[1]. Didn't end up being implemented in a real world scenario, but pigeons are extremely good at recognizing patterns without losing attentiveness over time. There's a pretty good book[2] on pigeon intelligence for anyone wanting to know more about their abilities. I've an amateur ornithologist (it's a hobby), but never been a huge rock pigeon fan. However, I respect their abilities and how much we underestimate them. [1] https://books.google.com/books?id=HxU-9UeDCI0C&pg=PA497&lpg=PA497#v=onepage&q&f=false https://books.google.com/books?id=HxU-9UeDCI0C&pg=PA497&lpg=... [2] http://www.amazon.com/Pigeons-Fascinating-Worlds-Revered-Reviled/dp/0802143288 http://www.amazon.com/Pigeons-Fascinating-Worlds-Revered-Rev...
- dutchbrit 11y ago
  Interesting links. People tend to underestimate pigeons (I use to race them when I was younger). Another interesting article: https://en.wikipedia.org/wiki/Project_Pigeon https://en.wikipedia.org/wiki/Project_Pigeon
  - will_pseudonym 11y ago
    Neat! My uncle used to race them, and as a kid, I would take the bands off of their ankles and into the clock-in box.
benkuykendall 11y ago
That's really cool! Here's a similar study that showed pigeons could "learn" the solution to the Monty Hall problem http://psycnet.apa.org/journals/com/124/1/1 http://psycnet.apa.org/journals/com/124/1/1 The whole point of this problem is that it is counter-intuitive for humans -- but I don't think birds naturally have intuition for this kind of problem, so they solve it without bias, purely from observed probabilities. I wonder if it's useful to think of pigeon learning like a machine learning algorithm -- after being shown examples of images that are classified according to some relatively simple but unknown boolean function, the pigeons can predict the value of that function on new examples.
- yareally 11y ago
  > I wonder if it's useful to think of pigeon learning like a machine learning algorithm -- after being shown examples of images that are classified according to some relatively simple but unknown boolean function, the pigeons can predict the value of that function on new examples. Not sure about pigeons unless the images were simple enough, but crows, jays or parrots might be able to do that. My concern with pigeons is they might not have enough of the abstract/critical thinking abilities the other 3 bird species have. I could be underestimating them, but I'd probably put my bets on crows figuring it out before a pigeon. Though crows, jays and parrots are also smart enough to get bored after a while, so they might not be up to the task New Caledonia Crows are smart enough to use "meta tools" where the tool they're using is only a step towards their end goal. They've been observed using smaller sticks to reach for a larger stick that will allow them to reach a piece of food. Crows (in general) also have awesome memories and will remember the faces of people who were mean or nice to them years later while also passing it on to their offspring (who have never seen this person before). In a lab study, jays were shown to cache food if they knew they won't find food early in the morning the next day. If they knew they have a meal waiting for them the next day, they would cache less or not at all. I know there's parrots that have been trained to open up locked boxes where certain randomly assigned locks are redundant. The parrots will ignore the redundant locks and only open the ones that are being used. http://www.washington.edu/news/2012/09/10/crows-react-to-threats-in-human-like-way/ http://www.washington.edu/news/2012/09/10/crows-react-to-thr... link to the paper on crow facial recognition: http://www.pnas.org/content/109/39/15912.full http://www.pnas.org/content/109/39/15912.full
cing 11y ago
From the paper: "...when given a different task—namely, classification of suspicious mammographic densities (masses)—the pigeons proved to be capable only of image memorization and were unable to successfully generalize when shown novel examples. " Sounds like overfitting, we need to L2 regularize these pigeons.
kafkaesq 11y ago
I've often felt like one of those pigeons -- pecking away for hours on end the sake of "differential food [rent, etc] reinforcement" -- at many of my programming jobs.
- finance-geek 11y ago
  I'm doing Caffe setups for learning on XRays...and I also feel like a pigeon going thru all the scripts. Now that's meta!
jbandela1 11y ago
This article has a completely misleading title. In reading the article, it looks like pigeons were able to identify breast cancer on pictures of pathology slides (ie done using a microscope). The article states that they were not able to identify tumors on mammograms (the study that actually uses x-rays). So no pigeons were not taught to diagnose breast cancer on X-Rays.
i336_ 11y ago
Presuming unconditionally for the sake of argument that this methodology actually works, I'm curious - would it be possible to make this type of thing self-sustaining? As in, the pigeons are rewarded (or not) based on their choices within only a few moments, given only the pigeons' input to the system. If pigeons were isolated and input was fed to them in parallel, then two or more inputs were taken to mean consensus and interpreted as evidential truth, would the distribution between individual pigeon responses mean that false positives would be evened out across the input distribution, or would such a technique result in progressive drift of accuracy? I'm guessing some form of drift would be quite likely, and that for a real-world system a firehose of already-known outcomes (go vs. no-go) would need to be fed to the pigeons to ensure accuracy. Taking it further, I wonder if the drift could be modelled and accounted for...

35 points by randomname2 11y ago

Direct link to paper: http://www.plosone.org/article/fetchObject.action?uri=info:doi/10.1371/journal.pone.0141357&representation=PDF http://www.plosone.org/article/fetchObject.action?uri=info:d...

yareally 11y ago

Back in the 60s, there was a study done with pigeons being able to inspect pills on an assembly line[1]. Didn't end up being implemented in a real world scenario, but pigeons are extremely good at recognizing patterns without losing attentiveness over time. There's a pretty good book[2] on pigeon intelligence for anyone wanting to know more about their abilities. I've an amateur ornithologist (it's a hobby), but never been a huge rock pigeon fan. However, I respect their abilities and how much we underestimate them. [1] https://books.google.com/books?id=HxU-9UeDCI0C&pg=PA497&lpg=PA497#v=onepage&q&f=false https://books.google.com/books?id=HxU-9UeDCI0C&pg=PA497&lpg=... [2] http://www.amazon.com/Pigeons-Fascinating-Worlds-Revered-Reviled/dp/0802143288 http://www.amazon.com/Pigeons-Fascinating-Worlds-Revered-Rev...

dutchbrit 11y ago
Interesting links. People tend to underestimate pigeons (I use to race them when I was younger). Another interesting article: https://en.wikipedia.org/wiki/Project_Pigeon https://en.wikipedia.org/wiki/Project_Pigeon
- will_pseudonym 11y ago
  Neat! My uncle used to race them, and as a kid, I would take the bands off of their ankles and into the clock-in box.

benkuykendall 11y ago

That's really cool! Here's a similar study that showed pigeons could "learn" the solution to the Monty Hall problem http://psycnet.apa.org/journals/com/124/1/1 http://psycnet.apa.org/journals/com/124/1/1 The whole point of this problem is that it is counter-intuitive for humans -- but I don't think birds naturally have intuition for this kind of problem, so they solve it without bias, purely from observed probabilities. I wonder if it's useful to think of pigeon learning like a machine learning algorithm -- after being shown examples of images that are classified according to some relatively simple but unknown boolean function, the pigeons can predict the value of that function on new examples.

yareally 11y ago
> I wonder if it's useful to think of pigeon learning like a machine learning algorithm -- after being shown examples of images that are classified according to some relatively simple but unknown boolean function, the pigeons can predict the value of that function on new examples. Not sure about pigeons unless the images were simple enough, but crows, jays or parrots might be able to do that. My concern with pigeons is they might not have enough of the abstract/critical thinking abilities the other 3 bird species have. I could be underestimating them, but I'd probably put my bets on crows figuring it out before a pigeon. Though crows, jays and parrots are also smart enough to get bored after a while, so they might not be up to the task New Caledonia Crows are smart enough to use "meta tools" where the tool they're using is only a step towards their end goal. They've been observed using smaller sticks to reach for a larger stick that will allow them to reach a piece of food. Crows (in general) also have awesome memories and will remember the faces of people who were mean or nice to them years later while also passing it on to their offspring (who have never seen this person before). In a lab study, jays were shown to cache food if they knew they won't find food early in the morning the next day. If they knew they have a meal waiting for them the next day, they would cache less or not at all. I know there's parrots that have been trained to open up locked boxes where certain randomly assigned locks are redundant. The parrots will ignore the redundant locks and only open the ones that are being used. http://www.washington.edu/news/2012/09/10/crows-react-to-threats-in-human-like-way/ http://www.washington.edu/news/2012/09/10/crows-react-to-thr... link to the paper on crow facial recognition: http://www.pnas.org/content/109/39/15912.full http://www.pnas.org/content/109/39/15912.full

cing 11y ago

From the paper: "...when given a different task—namely, classification of suspicious mammographic densities (masses)—the pigeons proved to be capable only of image memorization and were unable to successfully generalize when shown novel examples. " Sounds like overfitting, we need to L2 regularize these pigeons.

kafkaesq 11y ago

I've often felt like one of those pigeons -- pecking away for hours on end the sake of "differential food [rent, etc] reinforcement" -- at many of my programming jobs.

finance-geek 11y ago
I'm doing Caffe setups for learning on XRays...and I also feel like a pigeon going thru all the scripts. Now that's meta!

jbandela1 11y ago

This article has a completely misleading title. In reading the article, it looks like pigeons were able to identify breast cancer on pictures of pathology slides (ie done using a microscope). The article states that they were not able to identify tumors on mammograms (the study that actually uses x-rays). So no pigeons were not taught to diagnose breast cancer on X-Rays.

i336_ 11y ago

Presuming unconditionally for the sake of argument that this methodology actually works, I'm curious - would it be possible to make this type of thing self-sustaining? As in, the pigeons are rewarded (or not) based on their choices within only a few moments, given only the pigeons' input to the system. If pigeons were isolated and input was fed to them in parallel, then two or more inputs were taken to mean consensus and interpreted as evidential truth, would the distribution between individual pigeon responses mean that false positives would be evened out across the input distribution, or would such a technique result in progressive drift of accuracy? I'm guessing some form of drift would be quite likely, and that for a real-world system a firehose of already-known outcomes (go vs. no-go) would need to be fed to the pigeons to ensure accuracy. Taking it further, I wonder if the drift could be modelled and accounted for...