AI Learns Like an Animal by Mimicking the Brain's Two-Part Neurons

How does an animal's brain learn to connect a neutral cue, like a bell, with an important event, like food? A new computational model provides fresh insight by mimicking a fundamental process called stimulus substitution, where the brain's response to the bell gradually transforms to resemble its response to the food.

This recurrent neural network is not just another AI. It incorporates two powerful features, or 'inductive biases', observed in the cortex. The first is how the brain represents stimuli, but the second is architectural: it models pyramidal neurons as having two distinct compartments. This two-part structure is believed to be a fundamental unit for associative learning in the brain.

The result is a biologically plausible model that learns a wide array of associations with training times comparable to animal experiments. Crucially, it achieves this without needing parameter fine-tuning for each new task, a significant limitation of more common 'Hebbian' learning rules. This research highlights how the unique organisation of our neurons may provide a powerful evolutionary edge in learning.