A new study has used a type of machine learning called a neural network to reveal how different kinds of training can change how learning happens—both in machines, and in living brains.
"We can use these complex models to make specific predictions about the functions of the brain regions we're interested in," says Jack Bowler, PhD, postdoctoral fellow in neurobiology at University of Utah Health and first author of the study. "If you get at the most abstract level, it's a fairly good analogy for how we think the brain has to work."
As neural networks learn, they adopt patterns of activity that are strikingly similar to real-life neuron firing patterns in a part of the brain that's important for learning, the researchers found.
What's more, the type of training affects both the activity pattern, and the ability to solve a task, in similar ways between neural networks and living brains. For both computational and living models, starting by learning a more simple task prepares the learner to excel at complex tasks. On the other hand, poorly structured learning experiences can bias learners to make specific, predictable errors when things get more complicated.
The results are published in Nature Neuroscience.
Neural networks learn—and fail—like living creatures
Jim Heys, PhD, associate professor of neurobiology at U of U Health and senior author on the study, compares the process of scaling from simple to complex tasks to learning math.
"Everyone knows that when you're trying to learn calculus, you first learn order of operations, and then you learn algebra, and you learn trigonometry, and you build up these concepts systematically," Heys says. "But why that is the case?"
Researchers trained the neural networks on a relatively complex task—responding in a specific way only after presented with two stimuli of different durations (a "go trial"), and not responding when the two stimuli are the same length ("no go"). It's the virtual equivalent of a complex learning task the team had previously developed to study learning in a mouse model, in which mice are rewarded for responding to timed patterns of smells.
For a mouse, learning to respond only to the correct patterns is a complex task that requires multiple training steps, building from simple to complex. The same is true for a neural network, the researchers found. Training a simpler task first by only giving the network "go trials" made the network respond more accurately once it progressed to the full, complex task with both "go" and "no go" trials.
In contrast, when researchers trained the networks on the complex task without providing a simpler one first—essentially, jumping straight to calculus before learning algebra—the networks tended to make repeated, predictable errors.
They often responded too early to "go" trials, jumping the gun after the first long stimulus. Importantly, this is the exact same error that living animals make when trained improperly, which shows that the networks can be used to predict how real brains learn.
Neural networks predict patterns of real-life brain activity
The researchers then measured neuron-level brain activity in mice as they responded to trials, focusing on a specific brain region known to be involved in task learning. The patterns of neuron activity they observed looked very similar, on a large scale, to the patterns of activity the neural networks adopted during trials.
Appropriately trained neural networks, like appropriately trained mice, moved through a cyclic pattern of activity over the course of a single trial, ending the trial in about the same state they began. During "go" trials, both neural networks and mice adopted a different pattern of activity: while they still moved through a cycle, the activity swerved off its regular path during the time period that the mouse—or the network—needed to respond to the trial.
The similarities in activity and behavior between neural networks and real-life brains suggest that the networks can be especially useful as a first step to understanding human learning, Bowler says. Discoveries with computational models help generate useful, targeted questions about how we think, which can then be tested with animal models—helping scientists learn vital information while using the smallest number of research animals.
"By modeling up front, we were able to move much more quickly to useful hypotheses and reduce the number of animals we needed to test," Bowler says. "It helps us work with animals more respectfully."
The findings could ultimately guide development of better trainings that help people scale up their learning, the researchers say. By revealing how the brain learns when it's healthy, the team hopes that their work could also be useful for understanding what goes wrong in diseases that affect complex thinking.
"We're understanding fundamental principles of how the brain works," Heys says. "So when it breaks, like in diseases like Alzheimer's disease, which affect really high level complex cognitive functions first, we can understand how to fix it."
