A Chinese research team has revealed how brain neurons support flexible decision-making. According to a study recently published in Science, rule-related information and sensory information are compared within the same neuron, which itself takes part in cognitive computation.

Traditional theory holds that neurons are the basic structural and functional units of the nervous system. For a long time, most studies treated each neuron as a simple point-like unit that performs summation.

However, researches in brain science and artificial neural networks have found that higher cognitive functions such as flexible decision-making require the integration of massive contextual information and abstract rules, as well as dynamic adjustment of output. Such complexity is difficult to explain with the simple structure of point-like neurons.

Xu Ninglong, a researcher at the University of Chinese Academy of Sciences, and his team designed a series of experiments in which mice heard different sounds and had to infer hidden rules. Once the mice had learned the current rules, the team changed the rules to see how quickly the mice adapted to the new logic. The team used two-photon microscopy to observe neurons in the mice's auditory cortex.

A neuron is like a tree with many branches. The cell body is the main part of the neuron, while dendrites are branch-like extensions. The distant branches are called apical dendrites.

The experiments found a clear division of labor within the same neuron. Apical dendrites mainly encode the abstract rule inferred by the mouse, while the cell body more strongly represents external auditory sensory input. The two types of signals undergo nonlinear integration within the neuron, producing a decision output that drives behavioral choice.

The experiments demonstrated that a single neuron is itself a microcomputer and that the interactions between a neuron's dendrites and cell body can perform decision-related cognitive computation. Only after a hypothesis has been verified multiple times and confirmed to be reliable does the brain slowly strengthen neuronal synapses and store the rule permanently.

The team found that although AI can also use context and existing information to adjust its answers, its mechanisms differ greatly from those of the brain. Large models mainly develop capabilities through training on data, whereas the brain combines experiential learning, internal states, and real-time information processing. Even a single neuron has a rich internal computational structure.

The team also found that the brain can support flexible behavior while using limited energy. Its biological mechanisms are worth studying for insights that could help make AI systems more flexible and more efficient.