A new study suggests that biological neurons have more control over their environment than previously thought, which could be reproduced in artificial neural networks used in machine learning to enhance artificial intelligence (AI). The team at the Flatiron Institute in the USA has developed a new model of neurons that views them as small “controllers,” able to influence their environment based on collected information.
The researchers note that previous models of neurons are outdated and do not fully capture all the computational capabilities that real neurons possess. They believe that updating these models could lead to more powerful artificial neural networks that better imitate the capabilities of the human brain. Most modern artificial intelligence tools, like ChatGPT, rely on a computational model of a living neuron from the 1960s.
Artificial neural networks aim to mimic how the human brain processes information and makes decisions, albeit in a simpler form. They are composed of layers of nodes, with input nodes at the beginning, intermediate nodes for processing, and output nodes for sending results. However, current networks only allow information to pass through nodes in one direction, without the ability for nodes to influence the information received from higher nodes.
The new model treats neurons as tiny “controllers,” allowing them to control the status of others. This approach could potentially improve the performance and efficiency of many machine learning applications, according to the researchers. The model is inspired by large-scale brain circuits made up of many neurons, and suggests that neurons may be able to predict their inputs even if they cannot directly influence them.
Moving forward, the team plans to explore different types of neurons that may not fit into this new model, such as those in the retina that receive direct input from the visual environment. While these neurons may not control their inputs in the same way as deeper brain neurons, they may still use some of the same principles to predict their inputs.