A new computational model of the brain, based closely on its biology and physiology, not only learned a simple visual category learning task exactly as well as laboratory animals, but even made it possible to discover counterintuitive activity in a group of neurons that researchers working with animals performing the same task had not previously noticed in their data, according to a team of researchers from Dartmouth College at MIT and the State University of New York at Stony Brook.
Notably, the model made it possible to achieve these achievements without having to learn from data from animal experiments. Instead, it was built from the ground up to faithfully replicate how neurons connect into circuits and then communicate electrically and chemically across broader areas of the brain to create cognition and behavior. Then, when the research team asked the model to perform the same task it had previously performed with animals (looking at patterns of dots and deciding which of two broader categories they fit into), they obtained very similar neural activity and behavioral results, acquiring the skill with almost exactly the same irregular progression.
“It's basically creating new, simulated graphs of brain activity that are only later compared to laboratory animals. The fact that they match each other as strikingly as they do in real life is kind of shocking,” he says. Richard Grangerprofessor of psychology and brain sciences at Dartmouth and lead author of the new study Nature communication which describes the model.
The goal of creating the model and newer iterations developed since the paper was written is to not only offer insight into how the brain works, but also how it might work differently in disease and what interventions could correct these aberrations, adds the co-author Earl K. MillerPicower Professor at the Picower Institute for Learning and Memory at MIT. Miller, Granger and other members of the research team founded the company Neuroblox.ai to develop biotechnological applications of models. Co-author Lilianne R. Mujica-Parodi, professor of biomedical engineering at Stony Brook and principal investigator of the Neuroblox project, is the company's CEO.
“The idea is to create a platform for biomimetic brain modeling that will help us more efficiently discover, develop, and improve neurotherapeutics. For example, drug development and testing of their effectiveness can take place earlier in the process on our platform, before the risks and costs of clinical trials are involved,” says Miller, who is also a faculty member of MIT's Department of Brain and Cognitive Sciences.
Creation of a biomimetic model
Anand Pathak, a postdoc at Dartmouth, created a model that differs from many others in that it includes both fine details, such as how individual pairs of neurons connect to each other, and large-scale architecture, including the effects of neuromodulatory chemicals such as acetylcholine on information processing in different regions. Pathak and team iterated their designs multiple times to ensure they met various constraints observed in real brains, such as how neurons synchronize within broader rhythms. Many other models focus solely on small or large scale, but not both, he says.
“We didn't want to lose the tree and we didn't want to lose the forest,” Pathak says.
The metaphorical “trees,” called “primitives” in the study, are small circuits consisting of several neurons each that connect based on the electrical and chemical principles of real cells to perform basic computational functions. For example, in the cerebral cortex model, one primitive design involves excitatory neurons that receive input from the visual system through synaptic connections influenced by the neurotransmitter glutamate. These excitatory neurons then densely connect to inhibitory neurons, competing with each other to give them the signal to turn off other excitatory neurons – a winner-takes-all architecture found in real brains that regulates information processing.
On a larger scale, the model includes four brain regions needed for basic learning and memory tasks: the cerebral cortex, brainstem, striatum, and a “tonically active neuron” (TAN) structure that can introduce a little “noise” into the system via bursts of acetylcholine. For example, because the model was intended to categorize presented dot patterns, TAN initially provided some variability in the way the model acted on visual data so that the model could learn by examining different actions and their outcomes. As the model continued to learn, cortical and striatal circuits strengthened the connections that suppressed TAN, allowing the model to act on what it learned with increasing consistency.
As the model engaged in the learning task, real-world properties emerged, including dynamics that Miller commonly observed in his animal studies. As learning progressed, the cortex and striatum became increasingly synchronized in the “beta” frequency band of brain rhythms, and this increased synchronization correlated with the moments at which the model (and animals) made correct judgments about what it saw.
Revealing “incoherent” neurons
However, the model also presented researchers with a group of neurons – about 20 percent – whose activity seemed highly predictable in the event of an error. When these so-called “incoherent” neurons influenced the circuitry, the model misjudged the categories. Granger says the team initially thought it was a quirk of the model. But then they looked at real brain data that Miller's lab had collected while the animals performed the same task.
“Only then did we go back to the data we already had, confident that it couldn't be there because someone would say something about it, but it was there, but it was never noticed or analyzed,” he says.
Miller says these counterintuitive cells may have a purpose: It's good to know the rules for doing a task, but what happens if the rules change? Trying alternatives from time to time can allow the brain to encounter a newly emerging set of conditions. Indeed, a separate laboratory of the Picower Institute recently published evidence that humans and other animals sometimes do this.
Granger says that while the model described in the new paper exceeded the team's expectations, the team continues to evolve it to make it sophisticated enough to handle a greater variety of tasks and circumstances. For example, they added more regions and new neuromodulatory chemicals. They also began to test how interventions such as drugs affect its dynamics.
In addition to Granger, Miller, Pathak and Mujica-Parodi, the paper's other authors are Scott Brincat, Haris Organtzidis, Helmut Strey, Sageanne Senneff and Evan Antzoulatos.
Research support was provided by the Baszucki Brain Research Fund of the United States, the Office of Naval Research and the Freedom Together Foundation.
