recurrent network

A recurrent network is a group of neurons wired in loops, so that their signals flow not just forward but also back onto each other. Picture a room full of people who keep whispering to the neighbors who just whispered to them: a comment can travel around the circle and return to where it started. Instead of a one-way relay where a signal passes through and is gone, a recurrent network lets activity circle back, so the cells can keep talking to themselves long after the original nudge has faded.

This feedback wiring lets the network do two things a simple forward chain cannot. First, it can sustain activity — a brief input can set off a loop of firing that keeps going on its own, which is one way the brain seems to hold a thought, a phone number, or a plan in mind for a few seconds. Second, it can amplify or sharpen signals: each neuron feeds a little excitement back to its partners, so a faint pattern gets reinforced and stands out, while loops of inhibition can quiet competing patterns. The same machinery can also cause trouble — if the loops are too strong and nothing brakes them, runaway feedback can build into the synchronized storm of a seizure.

Recurrent loops appear almost everywhere in the brain, from the cortex to the hippocampus, and they are central to how circuits remember, decide, and generate rhythms. The idea also crossed into artificial intelligence: the recurrent neural networks used to process speech and text borrow this same trick of feeding their own output back as new input, so that what happened a moment ago shapes what the network does next.