AI Distinguishes Birds That Even Experts Can't (sciencemag.org) 11
Slashdot reader sciencehabit quote Science magazine: It's a fact of life for birders that some species are fiendishly difficult to tell apart — in particular, the sparrows and drab songbirds dubbed "little brown jobs." Distinguishing individuals is nearly impossible. Now, a computer program analyzing photos and videos has accomplished that feat. The advance promises to reveal new information on bird behaviors...
The tool, called a convolutional neural network, sifts through thousands of pictures to figure out which visual features can be used to classify a given image; it then uses that information to classify new images. Convolutional neural networks have already been used to identify various plant and animal species in the wild, including 48 kinds of African animals. They have even achieved a more complicated task for elephants and some primates: distinguishing between individuals of the same species. Team member André Ferreira, a Ph.D. student at the University of Montpellier, fed the neural network several thousand photos of 30 sociable weavers that had already been tagged... [W]hen given photos it hadn't seen before, the neural network correctly identified individual birds 90% of the time, they report this week in Methods in Ecology and Evolution. Behavioral ecologist Claire Doutrelant of CNRS, the French national research agency, says that's about the same accuracy as humans trying to spot color rings with binoculars.
Ferreira then tried the approach on two other bird species studied by Damien Farine, a behavioral ecologist at the Max Planck Institute of Animal Behavior. The tool was just as accurate...
The tool, called a convolutional neural network, sifts through thousands of pictures to figure out which visual features can be used to classify a given image; it then uses that information to classify new images. Convolutional neural networks have already been used to identify various plant and animal species in the wild, including 48 kinds of African animals. They have even achieved a more complicated task for elephants and some primates: distinguishing between individuals of the same species. Team member André Ferreira, a Ph.D. student at the University of Montpellier, fed the neural network several thousand photos of 30 sociable weavers that had already been tagged... [W]hen given photos it hadn't seen before, the neural network correctly identified individual birds 90% of the time, they report this week in Methods in Ecology and Evolution. Behavioral ecologist Claire Doutrelant of CNRS, the French national research agency, says that's about the same accuracy as humans trying to spot color rings with binoculars.
Ferreira then tried the approach on two other bird species studied by Damien Farine, a behavioral ecologist at the Max Planck Institute of Animal Behavior. The tool was just as accurate...
Gravity Error Detected? (independent.co.uk) 61
jd (Slashdot reader #1,658) writes: The large scale maps of the universe show something is seriously wrong with current models of gravity and dark matter. The universe simply isn't clumping right and, no, it's not the new improved formula. As you go from the early universe to the present day, gravity should cause things to clump in specific ways.
It isn't. Which means dark matter can't be cold and general relativity may have a problem.
They need more data to prove it's not just a freaky part of the universe they're looking at, which is being collected.
"The new results come from the Kilo-Degree Survey, or KiDS, which uses the European Southern Observatory's Very Large Telescope to map the distribution of matter across our universe," according to the Independent: So far, it has charted roughly 5% of the extragalactic sky, from an analysis of 31 million galaxies that are as much as 10 billion light years away... That allows researchers to build up a picture of all matter in the universe, of which some 90 per cent is invisible, made up of dark matter and tenuous gas.
It isn't. Which means dark matter can't be cold and general relativity may have a problem.
They need more data to prove it's not just a freaky part of the universe they're looking at, which is being collected.
"The new results come from the Kilo-Degree Survey, or KiDS, which uses the European Southern Observatory's Very Large Telescope to map the distribution of matter across our universe," according to the Independent: So far, it has charted roughly 5% of the extragalactic sky, from an analysis of 31 million galaxies that are as much as 10 billion light years away... That allows researchers to build up a picture of all matter in the universe, of which some 90 per cent is invisible, made up of dark matter and tenuous gas.
Please note that the error in AI identification of birds is 10%. By following the link in the gravity item, I learned that the error in homogeneity is also 10%. While birders find 10% an entirely acceptable error rate, the physicists find 10% unacceptable. I guess that while wearing my "physicist's hat", I understand the latter; but, while wearing my "engineer's hat" I'm finding that 10% is almost negligible. When I wear my "birder's hat", I say, "Wish I could be that good."
The problem with convolutional neural networks and indeed algorithms like ID3, which can generate a classifier (tree) from a dataset by data mining, is that they cannot generate explanations of how they work, which makes it difficult for us humans to understand them and thus assess their ability. I had one of my AI team, Regina Reppenhagen, implement ID3 to generate decision trees as classifiers and another student to generate natural language explanations from the decision trees.
However, ID3 is susceptible to noise in the training atasets and doesn't know where to set a breakoff point amongst the noise, so it can't assess its own reliability :-(
Posted by: Ole Phat Stu | August 03, 2020 at 06:28 AM
Stu--Don't make me go down in there to convolve matrices. The last time I played with convolution was in the Rapid Runway Repair program at Tyndall AFB FL in 1981-1983. I don't even understand the language of convolutional neural networks. *sigh* That said, at this point in time I just have to use a "black box" functional relationship - or does that even make sense in this application? You brilliant programming geeks are too high up for me to even see you.
Posted by: Cop Car | August 03, 2020 at 07:35 AM
I need that AI for bird identification of look-alike birds!
Posted by: bogie | August 09, 2020 at 03:31 AM
I would share, had I the AI.
Posted by: Cop Car | August 09, 2020 at 03:44 AM