How artificial intelligence learned to distinguish birds

by   |  VIEW 204

How artificial intelligence learned to distinguish birds

A study published in Methods in Ecology and Evolution shows how a well-designed artificial intelligence is able to identify individual specimens of a bird species both in the wild and in captivity, and with a percentage of errors that does not exceed 10% .

A discovery that will prove particularly useful for studying the behavior of birds in the wild. In the study, the researchers used a deep learning technology called convolutional neural networks, particularly effective for image classification and already successfully used for the recognition of individuals in species such as pigs and elephants.

In the world of ornithology, however, this is the first time, because until now no one had managed to use an artificial intelligence capable of recognizing birds better than man. The results lived up to expectations: the program is in fact able to identify diamond specimens with an accuracy of 90% in the case of wild birds, and 87% for those in captivity.

The results of the study

André Ferreira, a researcher at the French CNRS Center of Ecology Fonctionnelle et Evolutive and first author of the study, said about the technology: "We have shown that computers can reliably identify dozens of bird specimens, even though we humans do not succeed in any way to distinguish them from each other.

This will allow us to overcome one of the main problems that arise in the study of birds in the wild, namely that of identifying individual specimens." At the moment the program still has a significant limit: it is able to identify only the specimens contained in the dataset on which it was trained.

When a new bird arrives in the area, it loses its usefulness, and is unable to recognize it unless it is tagged with an electronic label, taking a photo of it and having the program re-trained. Among the other problems that had slowed down research in the field, one of the main ones was the difficulty in obtaining material with which to train the program.

To teach a computer how to recognize the content of an image, it is necessary to submit a set of photographs on which to train, so that, thanks to machine learning, you will learn by trial and error as a human would do. To overcome this problem, the researchers of the study used feeders equipped with photographic traps and sensors, placed in areas where many birds of the Taeniopygia guttata species had already been tagged with special electronic labels to study their behavior in wild and in captivity.