This A.I. Can Recognize Individual Birds of the Same Species
Humans can’t reliably tell birds of the same species apart, limiting our ability to study their behavior, but the new A.I. is 90 percent accurate
Imagine seeing a group of birds of the same species cavorting in a nearby tree. Now imagine trying to tell each individual bird apart. It might be possible to fix your eye, or your binoculars, to one particular feathered friend for a matter of minutes, or if you’re particularly keen, hours. But come back to the same tree the next day and you’d be utterly lost trying to pick out the bird you’d spent the previous day ogling, if it’s there at all.
The problem of identifying individual birds has bedeviled birders and researchers studying bird behavior for time immemorial, but now new artificial intelligence-powered software has bested human birders once and for all, reports Erik Stokstad for Science.
“We show that computers can consistently recognize dozens of individual birds, even though we cannot ourselves tell these individuals apart,” says André Ferreira, a Ph.D. student at the University of Montpellier and the new study’s lead author, in a statement. “In doing so, our study provides the means of overcoming one of the greatest limitations in the study of wild birds – reliably recognizing individuals.”
The new technique is similar to facial recognition software used by smartphones and social media companies in which the A.I. is “trained” on labeled photos in order to eventually recognize a face in unlabeled ones. But to train the A.I. to pick out individual birds, the researchers needed to get their hands on enough labeled photos of the flying critters, reports Michael Le Page for New Scientist.
“We need thousands of pictures of the same individual,” Ferreira tells New Scientist. “With humans, this is easy. With animals, it is hard to do.”
To get enough pictures of pre-labeled birds, Ferreira relied on a tried and true but time-consuming work around: colored leg bands. Tracking birds with these bands has significant drawbacks, including the stress of tagging the animals and the many hours of analyzing photos or videos back in the lab required to glean useful data, according to Science. To make things easier, the researchers also equipped the leg bands of a group of sociable weaver birds (Philetairus socius) with radio transponders that were set up to trigger remote cameras.
With a supply of well-labeled bird photos, Ferreira and his colleagues set about training the machine learning algorithm, called a convolutional neural network, on thousands of images of the 30 sociable weavers, which as their name suggests, tend to hang out in large groups and weave complex, communal nests.
The researchers trained the system to recognize the wild sociable weavers as well as captive zebra finches and wild great tits and found it was roughly 90 percent accurate when presented with a single image, the researchers reported last week in the journal Methods in Ecology and Evolution.
The technique shows promise but Ferreira and other researchers say that right now it still has significant limitations. For researchers studying elusive or endangered species, the necessity of tagging the animals and then obtaining large numbers of photos for training the A.I. may not be feasible, according to Science.
The system may also balk if a bird’s appearance changes significantly over the course of its life or even from one month to the next during seasonal molts. But the system’s most fundamental limitation is its need to be taught what a bird looks like before it can identify it.
“The model is able to identify birds from new pictures as long as the birds in those pictures are previously known to the models. This means that if new birds join the study population the computer will not be able to identify them,” says Ferreira in the statement.
The team hopes these last two limitations can be overcome through tweaks to the algorithm as well as even larger supplies of photos, spanning long periods of time. In the statement, the researchers say they’re currently at work on this larger task.