3 ms·
I don't remember if it was Skinner or someone else, but another great example of using pigeons was inspecting parts in a factory. The automated line was suppose
by tzs 4y ago
I don't remember if it was Skinner or someone else, but another great example of using pigeons was inspecting parts in a factory. The automated line was supposed to be producing identical parts but occasional one have a defect that could easily be spotted visually.
They had humans whose job was to just sit there watching the stream of parts come off the line and if they saw a defective one pull it off the line.
The problem was that humans were great at that for a short time after their shift started, and then their accuracy at noticing bad parts would drop way off.
A researcher proposed trying pigeons. They trained a pigeon to press a button when it was show a defective part and not press when shown a part that was not defective.
The pigeon was more accurate than humans and it could go all day without losing accuracy.
But they couldn't just deploy the pigeon because it needed to occasionally be rewarded for correctly identifying defective parts to reinforce the behavior. Otherwise over time he would revert to not caring about the parts and buttons.
The solution to that was to deploy pigeons in groups of 3. The 3 would all see the same part at the same time. Their buttons were wired to a system that would:
1. if at least 2 out of 3 press their buttons trigger something that kicks the current part off the line, and
2. if 2 pressed their button and the other one did not would trigger automated food dispensers to give those 2 a reward.
The idea with #2 is that as the learning starts to fade the number of 2 of 3 pressing the button would go up, which would reward the two that got it right. This provides intermittent rewards that would sustain the learning.
To train new pigeons they extended the system so that a 4th pigeon could be added to a group of 3. The 4th pigeon's presses would not count toward deciding the fate of the current part, but the 4th pigeon would be rewarded for a press if a majority of the group of 3 pressed.
This let them have a group of 3 already trained pigeons train up a new pigeon with minimal human intervention.
Technically the pigeon based inspections were a complete success. Much cheaper than humans, much more accurate, and the accuracy stayed high throughout their shift.
At the conclusion of test though the company declined to go forward and put the pigeons into production. Their worry was that the public would have a hard time accepting pigeon inspected parts no matter how much research they could cite showing the pigeons were superior in pretty much every way than humans at that job.
PS: I've probably got some details wrong. I heard about this in a psychology class I took around 40+ years ago.
- saghm 4y ago> if at least 2 out of 3 press their buttons trigger something that kicks the current part off the line If wherever they kicked the part off to wasn't called a "pigeon hole", somebody wasn't doing their job
- noipv4 4y agovery interesting! learning using neural networks of pigeon. Do you have a link to this experiment? I would like to learn more about how pigeons learn and inspect.
- fblp 4y agoBest hn comment I've read in a while!
- tibbydudeza 4y agoHumans will flake out when doing boring stuff - pigeons are just stupid than the usual biological drives of where the next meal will come from. Intelligence is sometimes a curse. Also, they don't have the concept of time scale - one day is the same to them as the next one.
- kqr 4y agoI'm curious what the accuracy would be of three humans that are given positive feedback, to give the pigeons more of a fair fight. (More expensive, of course.)