These experiments inspired Elizabeth Buffalo, a learning and memory researcher at the University of Washington, to develop a virtual foraging task in monkeys that could probe learning and memory without relying on overtrained behaviors. “We’re interested in episodic memory formation that’s different from overtrained tasks,” Buffalo says. “Our hope is that in VR we can make things naturalistic enough to then be able to capture experiential learning built on a framework that we can control.”
Buffalo was specifically interested in understanding whether the hippocampus encodes memories the same way in primates as it does in rodents. What her team found surprised her. In a large 3D foraging environment, she couldn’t easily identify the place cells characteristic of the rodent brain. Instead, the cells in the hippocampus seemed to encode the different stages of the task that the animals were performing. That insight was made possible by the combination of large-scale recordings during a naturalistic task.
“Having the large-scale recordings was really critical for seeing this kind of representation of the tasks across the population,” Buffalo says. “If you were just looking at the data neuron by neuron, you might say that this structure has a really complicated mixture of responses: place, gaze, reward, delay, etc. But if you instead take a look at it at the response of the population, what you see is that the population activity really carves out behaviorally relevant epochs in the task.”
With more complex tasks and plentiful neurons, researchers are also able to identify differences between individual animals. Like humans, laboratory animals have different personalities — one animal might be risk averse, while others might be prone to taking chances. With enough data from each individual, researchers can examine the neural mechanisms underlying individual differences in behavior. “There’s a big opportunity to start going after individual-difference neuroscience, which is a big and important thing,” Hayden says. “This is just an example of the kind of thing that gets uncovered when you make things more naturalistic, and you move away from these highly controlled, simplistic tasks.” Hayden’s parallel work recording neural activity in humans is starting to demonstrate that different brains may compute tasks differently, a finding that has major implications for the way we approach neurological disorders.
Tackling the complexity of rich naturalistic datasets
Researchers analyzing stereotyped tasks, where an animal chooses between two or three options, have an established toolbox to draw on. They can align their neural data by specific events that occur across trials, such as the animal’s lever press or the delivery of a reward, and average the data to derive an event-evoked neural signal. But without the familiar trial structure of a two-choice task, the analysis landscape for naturalistic tasks looks a little bit like the wild west. Researchers have applied behavioral economics and ecological frameworks to foraging, but many of these models do not generalize well to more complicated foraging tasks. Moreover, researchers are collecting data from more and more neurons simultaneously, adding another layer of complexity to the task of understanding the experiment as a whole. As a result, many researchers are turning to both old and new theoretical frameworks to make sense of the data.
“The most challenging thing in these experiments is actually not the [recording] technology, but the data analysis,” Angelaki says. “We cannot take off-the-shelf analyses. I need to collaborate with data scientists because we really need to develop our own algorithms.” Her lab, in collaboration with Cristina Savin at NYU, is adapting algorithms such as canonical correlation analysis and demixed principal component analysis to use with their neural foraging data.
Angelaki and other experimentalists are also reaching out to theorists who can use computational modeling to make sense of the data. Buffalo is teaming up with Adrienne Fairhall and Xiao-Jing Wang, both SCGB investigators, while Angelaki is collaborating with Pitkow, Savin and Paul Schrater of the University of Minnesota to test reinforcement learning frameworks. Reinforcement learning models, in combination with deep learning, can be used to understand the kind of mental models that animals are developing while doing the task, and can in turn help us understand the brain.
During a task, experimenters have access to external variables, such as the animal’s movement and choices. However, they do not have access to the animal’s thought processes or its evolving understanding of its environment. Behavioral modeling helps researchers identify these hidden, or latent, variables and use them to explain the neural activity — Pitkow dubs this a sort of “mathematical mind reading.”
A good mental model for the animal is helpful because then we can impute these latent variables — an estimate of the animal’s internal beliefs — as a function of time, Pitkow says. We can then check if the latent variables calculated from the behavioral model match what’s going on in the brain, and specifically whether the evolution of neural activity is consistent with the evolution of the latent variables, he says. Using foraging tasks lends the ability to build a much richer model of the behavior and ultimately distill latent variables that may not even be present in a simple alternative forced-choice task.
Hayden, Angelaki, Pitkow and others are also digging out old engineering textbooks and thinking about their data in terms of control theory, a dynamical systems model that situates the brain as a ‘controller’ and can help quantify the importance of feedback loops and latent variables.
Other researchers, like Sue Ann Koay, who started her lab last year at the Howard Hughes Medical Institute’s Janelia Research Campus in Virginia, are thinking up radical new ways to devise hypotheses about brain function. In Koay’s randomly generated virtual environments, mice do not know the configuration or even the extent of the world they are in, so there is not a well-defined way to complete the task. Mice can find food in unexpected places in a world that may never end and may sometimes even distort the normal geometry of space, as in a science fiction movie.
Many ecological models optimize for an efficient solution to foraging tasks by balancing time and resource trade-offs. Koay wants to explore models that are more biological in nature. She plans to use these models to inspire hypotheses about the ways brains can work, employing ideas from evolution to constrain her models. “I would like to have some method of exploring brain-like solutions that is not tied to a human-formulated list of tasks that they should solve,” she says. “The only thing I could think of that is more general is evolution.”
Koay is part of a broader community of young researchers who are thinking about foraging in new ways, developing novel methods to generate hypotheses about how the brain facilitates foraging. Other early-career researchers are working on ways to bring ethology back into the conversation, hoping that neuroscientists can learn from their wealth of knowledge.
Last year, a team of trainees from both Janelia and the Champalimaud Centre for the Unknown in Lisbon kicked off the Future of Foraging seminar series, which features neuroscientists like Hayden as well as ecologists and ethologists who have been thinking about the foraging brain since long before we were able to record thousands of neurons simultaneously. Laura Grima, one of the co-founders of the series and a postdoctoral fellow at Janelia, is hopeful that a bridge between ethology and cutting-edge neuroscience is possible. “It is an overwhelming amount of data, even when you scale up some of these aspects just a bit, even if you don’t go the whole hog and just have mice running around in the wild,” Grima says. “I just hope as a field, we start to not be so scared of complexity and instead start to embrace it.”