Summary: A biologically based computer model built to mimic real emotional states, untrained on animal data, learned the classification task as real lab animals did, matching their accuracy, variability, and underlying emotional rhythms. By combining fine-scale synaptic rules with large structures throughout the cortex, striatum, brainstem, and acetylcholine-modulated systems, the model also produced characteristic learning patterns, including beta-band synchronization between regions during appropriate decision-making.
It also revealed a set of “wrong neurons” that predicted errors, a signal researchers only noticed in their animal data after the model revealed it. This biomimetic platform provides a powerful blueprint for assessing disease-related regional changes and evaluating therapeutic interventions in silico, offering a new way to develop next-generation neurotherapeutics.
Important Facts
- Biology-First Design: The model embeds actual neuronal communication rules, neurotransmitter dynamics, and multi-circuit structures to replicate biological computations.
- An urgent fact: It produced learning patterns, beta synchrony, and decision patterns similar to lab animals—even without training on biological datasets.
- Hidden Signs Revealed: The discovery of “wrong neurons” reveals an overlooked error prediction function that exists in the real brain.
Source: Picower Institute at MIT
A new computer model of the brain based on its biology and natural sciences not only studied the simple learning task of the visual class and laboratory animals, but even enabled the discovery of an opposing activity by a group of neurons researchers working with animals to perform the same activity that they had not noticed in their data before, said a team of scientists at Dartmouth College, Brook New York, State University of Stony University.
Remarkably, the model produced these achievements without being trained on any data from animal experiments. Instead, it was built from scratch to faithfully represent how neurons connect in circuits and communicate electrically and chemically across brain regions to produce cognition and behavior.
Then, when the research team asked the model to do the same task they had done with the animals before (looking at the dot patterns and deciding which of the two broad fields they fit), it produced an emotional activity that closely resembled the behavioral results, achieving an ability almost exactly the same as the adaptive progression.
“It produces new simulated episodes of brain activity that are then compared to lab animals. The fact that they are so strikingly similar is shocking,” said Richard Granger, professor of Psychological and Brain Sciences at Dartmouth and senior author of the new study in Natural Communication that defines the model.
The goal of the model, with new iterations developed since the paper was written, is not only to provide insight into how the brain works, but also how it can work differently in diseases and what interventions can correct those deviations, added co-author Earl K. Miller, the Picower 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 biotech applications for the models. Co-author Lilianne R. Mujica-Parodi, a professor of biomedical engineering at Stony Brook who is the Neuroblox Project's Principal Investigator, is the company's CEO.
“The idea is to create a platform for biomimetic modeling of the brain in order to have a more efficient way to discover, develop and improve neurotherapeutics. Drug development and efficacy testing, for example, can be done early in the process, in our environment, before the risk and cost of clinical trials.” said Miller, who is also a faculty member in MIT's Brain and Cognitive Sciences department.
Biomimetic modeling
Dartmouth postdoc Anand Pathak created the model, which is different from many others because it includes both small details, such as how individual pairs of neurons communicate, and large structures, including how information processing in all regions is affected by neuromodulatory chemicals such as acetylcholine.
Pathak and team replicated their designs to ensure they obeyed various constraints seen in the real brain, such as how neurons are synchronized to broad rhythms. Many other brands focus only on small or large scale but not both, he said.
“We didn't want to lose the tree, and we didn't want to lose the forest,” Pathak said.
Prototypical “trees”, called “primitives” in research, are small circuits of a few neurons that each connect based on the electrical and chemical principles of real cells to perform basic computational tasks.
For example, within a version of the brain cortex model, one classic design has excitatory neurons that receive input from the visual system through synapse connections mediated by the neurotransmitter glutamate.
Those excitatory neurons then make more connections with inhibitory neurons in a competition to signal them to shut down other excitatory neurons—“winner-takes-all” structures found in the real brain that control information processing.
On a large scale, the model includes four brain regions necessary for basic learning and memory functions: the cortex, the brainstem, the striatum and the structure of the “tonically active neuron” (TAN) that can add a little “noise” to the system with a burst of aceytlcholine.
For example, as the model performs the task of categorizing presented patterns of dots, TAN initially ensures some variation in how the model processes visual input so that the model learns by testing different actions and their consequences.
As the model continued to learn, circuits in the cortex and striatum strengthened connections that suppressed the TAN, causing the model to act on what it was learning with increasing consistency.
As the model performed the learning task, real-world structures emerged, including variables that Miller had observed often in his research on animals. As learning progressed, the cortex and striatum became more synchronized in the “beta” frequency band of the brain's rhythm, and this increase in synchronization correlated with the times when the model (and the animals) made the correct category of judgment about what they saw.
Reveals 'discordant' neurons
But the model also introduced the researchers to a group of neurons—about 20 percent—whose activity appeared to predict a major error. When these so-called “incongruent” neurons influence circuits, the model can make an incorrect class of judgment. At first, Granger said, the team realized it was a quirk of the model. But then they looked at real brain data Miller's lab collected when animals performed the same task.
“Until we went back to the data we already had, we were sure that this would not have happened because someone would have said something about it, but it was there and it had never been noticed or analyzed,” he said.
Miller said these opposition cells may serve a purpose: It's all well and good to learn the rules of the job, but what if the rules change? Trying alternatives from time to time can enable the mind to catch up with a new set of situations. Indeed, a separate Picower Institute lab recently published evidence that humans and other animals sometimes do this.
While the model described in the new paper exceeded the team's expectations, Granger said, the team was increasing its complexity enough to handle a variety of tasks and situations. For example, they added more regions and new neuromodulatory chemicals. They have also begun to test how interventions such as drugs affect its effectiveness.
In addition to Granger, Miller, Pathak and Mujica-Parodi, other authors of the paper are Scott Brincat, Haris Organtzidis, Helmut Strey, and Evan Antzoulatos.
Sponsorship: The Baszucki Brain Research Fund, United States, Office of Naval Research, and the Freedom Together Foundation provided research support.
Important Questions Answered:
A: It learned a visual category task with nearly identical patterns of progress, neural activity, and learning ability—even without training on biological data.
A: It revealed a number of “mismatched neurons” that predict errors. When the researchers examined data from older animals, a similar pattern was present but not observed.
A: It provides a platform for evaluating brain computation, simulating disease states, and evaluating neurotherapeutics before moving to risky and expensive trials.
Editor's Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper is fully revised.
- Additional content added by our staff.
About this AI, learning, and neuroscience research news
Author: David Orenstein
Source: Picower Institute at MIT
Contact person: David Orenstein – Picower Center at MIT
Image: Image posted in Neuroscience News
Actual research: Open access.
“Biomimetic modeling of corticostriatal micro-assemblies discovers the neural code” by Richard Granger et al. Natural Communication
Abstract
Biomimetic modeling of corticostriatal micro-assemblies discovers the neural code
Although mathematical models have deepened our understanding of neuroscience, it is still a great challenge to link real low-level physical activity (spiking, field strength) and biochemistry (transmitters and receptors) directly with high-level cognitive abilities (decision-making, working memory) and related disorders.
Here, we present a multi-scale model that is accurate through the use of machinery that generates modified physiology in which extended emotional and cognitive states emerge.
The model generates spiking, fields, phase synchronization, and synaptic changes, which directly generate working memory, decisions, and categorization.
These were then validated on extensive macaque experimental data where the model received no prior training of any kind. In addition, the simulation revealed a previously unknown neural code (“incoming neurons”) that directly predicts future errant behavior, which was also confirmed in the empirical data.
Thus the biomimetic model directly and predictably links decisions and reinforcement signals, computational interest, and spiking codes and field, neurobiological significance.
