New computer model of real neurons could lead to better AI

Almost all of the neural networks that power modern artificial intelligence tools like ChatGPT are based on a 1960s computer model of a living neuron. A new model developed at the Flatiron Institute’s Center for Computational Neuroscience (CCN) suggests that this decades-old approximation does not capture all of the computational capabilities that real neurons possess and that this old model is potentially holding back the development of AI.

The new model developed at CCN posits that individual neurons exert more control over their environment than previously thought. The updated neural model could ultimately lead to more powerful artificial neural networks that can better capture the powers of our brains, the model’s developers say.

The researchers present the model in an article published in the journal Proceedings of the National Academy of Sciences.

“Neuroscience has progressed a lot over the past 60 years and we now recognize that previous neuron models were rather rudimentary,” says Dmitri Chklovskii, group leader at CCN and lead author of the new paper. “A neuron is a much more complex and much more intelligent device than this oversimplified model.”

Artificial neural networks aim to mimic the way the human brain processes information and makes decisions, albeit in a much more simplified way. These networks are made up of ordered layers of “nodes” based on the neural model of the 1960s. The network begins with an input layer of nodes that receives information, then has intermediate layers of nodes that process the information, and then ends with an output layer of nodes that sends the results.

Typically, a node will only pass information to the next layer if the total input it receives from the previous layer’s nodes exceeds a certain threshold. When today’s artificial neural networks are trained, information flows through a node in only one direction, and nodes have no way to influence the information they receive from nodes earlier in the chain.

In contrast, the recently published model treats neurons as tiny “controllers,” a technical term for devices that can influence their environment based on information gathered about that environment. Our brain cells are not simple passive input relays, but can actually function to control the state of their fellow neurons.

Chklovskii believes that this more realistic model of a neuron as a controller could be an important step toward improving the performance and efficiency of many machine learning applications.

“Even though the achievements of AI are very impressive, many problems remain,” he says. “Current apps can give you wrong answers or even hallucinate, and training them takes a lot of energy; they’re very expensive. There are all these problems that the human brain seems to avoid. If we were to understand how the brain actually works By doing this, we could build better AI.”

The neuron-as-controller model was inspired by what scientists understand about the large-scale circuits of the brain made up of many neurons. Most brain circuits are thought to be organized in feedback loops, in which cells later in the processing chain influence what happens earlier in the chain. Much like a thermostat maintaining the temperature of a home or building, brain circuits must remain stable to avoid overloading the body’s system with activity.

Chklovskii says it wasn’t entirely intuitive that this type of feedback control could also be achieved by an individual brain cell. He and his colleagues realized that a new form of control, known as data-driven direct control, is simple and effective enough to be biologically plausible and take place in individual cells.

“People thought that the brain as a whole, or even parts of the brain, was a controller, but no one suggested that a single neuron could do that,” says Chklovskii. “Control is a computationally intensive task. It is difficult to think that a neuron has sufficient computational capacity.”

Viewing neurons as mini-controllers also explains several previously unexplained biological phenomena, says Chklovskii. For example, it has long been known that there is a lot of noise in the brain, and the purpose of this biological randomness has been debated, but the SCC team discovered through their modeling that certain types of noise could reality improve the performance of neurons.

Specifically, at junctions where one neuron connects to another (called “synapses”), it is often the case that one neuron transmits an electrical signal but the downstream coupled neuron does not receive the message. Whether and when the downstream neuron receives a synaptic signal appears to be largely governed by chance.

While other scientists theorized that such randomness was simply the nature of small biological systems and was not important for the behavior of neurons, the Flatiron team found that adding noise to their model neuron allowed it to adapt to a constantly changing environment. Randomness appears to be important for replicating how real neurons work, the team found.

Chklovskii next plans to analyze the types of neurons that don’t fit their new model. For example, neurons in the retina receive direct information from the visual environment. These neurons may not be able to control their inputs the way neurons deeper in the brain can, but they might use some of the same principles identified by Chklovskii and his team: namely, these neurons might be able to predict their entries, even if they can. I don’t influence them.

“Control and prediction are actually very related,” says Chklovskii. “You cannot control effectively without predicting the impact of your actions on the world.”

Provided by the Simons Foundation