Neuron bursts can imitate well-known artificial intelligence learning strategies

But in order for this teaching signal to solve the credit allocation problem without “pausing” in sensory processing, their model needs another key part. The team of Naud and Richards proposed that neurons have different compartments on the top and bottom, which process neural codes in completely different ways.

“[Our model] It shows that you can really have two signals, one rising and the other falling, and they can pass each other,” Naud said.

To make this possible, their model assumes that the tree-like branch that receives input at the top of the neuron listens only for bursts—internal teaching signals—to adjust their connections and reduce errors. The adjustment is done from top to bottom, just like in backpropagation, because in their model, the neurons at the top are adjusting the likelihood of the neurons below them sending impulses. Researchers have shown that when the network has more bursts, neurons tend to increase the strength of their connections, and when the bursts are less frequent, the strength of the connections tends to decrease. The idea is that the burst signal tells the neurons that they should be active during the task, strengthens their connections, and if doing so can reduce errors. No burst tells the neurons that they should be inactive and may need to weaken their connections.

At the same time, the branch at the bottom of the neuron treats the burst as a single spike-a normal external world signal-which allows them to continue sending sensory information up the circuit without interruption.

“In retrospect, the proposed idea seems logical, and I think it illustrates its beauty,” said Sacrament of john, A computational neuroscientist at the University of Zurich and ETH Zurich. “I think this is great.”

Others have tried to follow a similar logic in the past. Twenty years ago, Conrad Colding University of Pennsylvania and Peterkin University of Osnabrück, Germany suggested A learning framework with two-compartment neurons. But their proposal lacks many specific details related to biology in the new model, and it is just a proposal-they cannot prove that it can indeed solve the credit allocation problem.

“At the time, we just lacked the ability to test these ideas,” Kolding said. He considers this new paper to be “a great job” and will follow up in his own laboratory.

With today’s computing power, Naud, Richards and their collaborators successfully simulated their model, in which bursting neurons played the role of learning rules. They show that it solves the credit allocation problem in a classic task called XOR, which requires learning to respond when one of the two inputs (but not both) is 1. They also showed that the deep neural network built with their burst rules can approximate the performance of the backpropagation algorithm on challenging image classification tasks. But there is still room for improvement, because the backpropagation algorithm is still more accurate, and neither can fully match human capabilities.

“There must be details that we don’t have, and we must make the model better,” Naud said. “The main goal of this paper is to say that the kind of learning that a machine is doing can be approximated by a physiological process.”