Newswise — What do animals spend their time doing? To answer this question, researchers traditionally spent hours in the field closely observing a focal animal and classifying its movements into a set of behaviors: grooming, foraging, and the like. But what about animals that aren’t easily observed? More importantly, what if you don’t enjoy waking up at 3am to intently watch a goose while shivering in the snow?
Drs. Natnael Hamda and Andrew Hein at NOAA’s Southwest Fisheries Science Center and the University of California, Santa Cruz are merging behavioral ecology with computer science to fix these problems. Their efforts are bringing animal behavior into the 21st century world of big data.
“Ecology and behavioral science have gone through this transition of having almost no data, to having an enormous, really a flood of data – and in many ways, we aren’t ready for that,” says Hein.
The surge of data follows the rise in cheap, automated methods for recording animal movements. For example, miniature GPS tags can automatically track an animal’s speed or position, while precise cameras can capture hours of complex animal movements during behavioral trials in the lab.
Such datasets, often from hundreds of animals, can be powerful, but traditional tools fall short in making sense of them. That’s where the field of machine learning comes in.
In this genre of computer science, machines are allowed to “learn” patterns in datasets and make predictions about future events without being explicitly programmed. As Hamda puts it, “It’s unsupervised, there is no human interference. Can we get something out of that?”
The NOAA team is using machine learning to analyze a dataset from hundreds of tagged juvenile salmon in the Sacramento-San Joaquin River Delta, California. The data were originally collected by the California Department of Water Resources in order to find out why river salmon weren’t making it to the ocean during migration. Hein notes that the problem of failed oceanward migration appears to be one of the key challenges facing imperiled salmon populations.
Invasive predatory fish like striped bass were also tagged. By using machine learning to characterize salmon and predator behavior, the team can identify predation events. If a salmon suddenly starts moving like a predator, the salmon and its tag were likely eaten. By characterizing the timing and location of predation events, management decisions can be made to help salmon avoid getting eaten.
But getting machines to learn isn’t easy. Machine learning has mainly been applied to other types of data, like financial records or cell phone logs. There isn’t a cookbook in place to easily convert the methods for analyzing behavioral data. Hein likens the relatively young machine learning literature to still being somewhat “like the wild west”, with Hamda also noting that, “Choosing the right technique can be tricky”.
But the payoff is worth it. The technique can be applied to hundreds of animals at the same time, including those like fish that aren’t easily observed. “It’s also really good at finding things our eyes don’t pick up easily,” says Hein, which means that new behaviors can be identified that have been missed by our limited and biased human observations.
Ultimately, making biological sense of a computer generated set of behaviors may be one of the emerging field’s biggest challenges. A recent study used machine learning to document over 100 behavioral states in fruit flies, although it is unclear what many of these behaviors mean biologically.
Hamda and Hein are quick to point out the importance of coupling machine learning techniques with more traditional behavioral observations, noting that “Ultimately, without doing that, you just have a list of numbers.” So while big data and big computers might be the future of behavioral ecology, someone will still have to get up early and shiver with the geese in order to put names to those numbers.