Scientists in China say a single brain cell can behave like a tiny computer, weighing evidence and testing guesses on the spot. The finding, from the Chinese Academy of Sciences' Center for Excellence in Brain Science and Intelligence Technology, could point the way toward a more flexible generation of artificial intelligence (AI).

Today's AI models outmatch humans at language and image recognition, but only after training on thousands of times more data. Brains learn far faster, and researchers suspect the answer lies in how individual neurons handle information.

Neurons link up through synapses, the equivalent of roads between stations, and learning strengthens those roads. But a neuron is not just a junction. It grows branch-like dendrites, and its longest, farthest-reaching branch, the apical dendrite, carries abstract rules the brain has inferred. The cell body, meanwhile, receives raw sensory signals from the ears or eyes. Both streams meet inside a single cell, letting it compare a prediction with reality immediately.

Mice, sounds and rules

To test the idea, the team trained mice to spot patterns in sounds, then quietly swapped the patterns to see how fast the animals adapted. Using two-photon microscopy to watch neurons in the auditory cortex, the researchers saw dendrites and cell bodies carry out the comparison themselves, suggesting a lone neuron performs the computation.

Only once a guess proves reliable does the brain reinforce the synapse and store the rule for good.

Current AI imitates only the synapses, the roads between cells. If engineers can incorporate the newly found computing happening inside neurons, they may build systems that form hypotheses and infer new rules from just a few examples.

Evolution, the team notes, squeezed precise computation into every single neuron.

This might only be a glimpse into the tip of that iceberg, said a CMG report on the finding.