Picture an aquarium with a hundred fish from the same species swimming around. Now try to track one of them without losing sight. Tricky, right? What if you closed your eyes for a few seconds? Then it would be nigh on impossible. But that was the starting point for a group of researchers from Collective Behavior Lab at the Champalimaud Centre for the Unknown (CCU) in Lisbon lead by the Spanish researcher Gonzalo de Polavieja. The team tried to establish if an AI system would be able to achieve such a feat. After all, de Polavieja had already tried his luck in 2014 by using conventional algorithms and the results of this technological project were far from encouraging, they barely managed to track a dozen of them. And, to cut a long story short, in the end, they did succeed. The software they have now developed, based on deep learning neural networks, can track a fish among a hundred with a 99.9% precision. Codenamed Idtracker.ai, it is available for download as open source code.
“A new AI-based software can track a single fish among a hundred with a 99.9% precision.”
The researchers first installed an aquarium with dozens of zebrafish and a camera above to track the movement of the specimens. The tests were carried out with thirty, fifty and a hundred fishes. Although a comparatively modest number, conventional software would require years to identify their movement patterns. And, as pointed out at the beginning of this article, would prove completely unfeasible for a human mind. This is because the complexity of the interactions in a system grows exponentially with each new addition.
Thus, once the researchers had given up on using conventional algorithms, they implemented two deep learning neural networks that replicate a brain and can learn on their own. The first one is used to discriminate the fishes from other elements in the environment, while the second one tracks each individual fish, with names such as George or Tom. Following this process, if there are any remaining doubts due to overlapping trajectories or mix-ups, conventional algorithms are used to clear them. After an hour processing the video feed, 99.9 % of the specimens can be identified.
Once the process is completed, the Artificial Intelligence software has already learned who is who in the fish shoal and can recognize any specimen by viewing a random fragment of video, whether George, Tom or any of the other 99 specimens. The system seems quite scalable and tests have already been made with up to 150 fish with extremely low error margins.
And what interest could there be in identifying a zebrafish? The answer lies in collective behaviors. A system based on the Idtracker.ai software could have urban security and safety applications, allowing to track an individual or study the behavior of crowds in different situations such as a supermarket or a concert. In the same way, it could establish collaborative processes between people, with potential applications in sociology.