A melody plays, and the brain doesn't just listen note by note. It works like a conductor who already knows the score: each next note is expected because the previous ones still echo in memory. But how do neurons hold on to that echo? A new paper proposes a model of a spiking neural network called PCL+—a network where signals are transmitted in short bursts, just like in the real brain. It learns without a teacher, simply by observing sequences.
The secret lies in connections with delays. A runner passes the baton: he must reach his teammate at the exact moment the teammate is ready to receive it. Same here: neurons strengthen connections so that the signal from a past event arrives just in time to predict the future. It's like an echo in the mountains: your shout returns with a delay and overlaps with a new sound, creating a sense of continuity.
The most interesting part is that the network learned to "fill in" what was missing. In a gesture task where some hand movements were hidden, PCL+ guessed them better than models without memory. The brain does the same: when we see a person winding up for a throw, we know a throw is coming, even if we look away for a moment.
Some neurons in the brain send signals with delays of up to 100 milliseconds — about the same time it takes a fly to dodge a flyswatter.
In Asimov's novel "Foundation," a scientist predicts the future of civilization using mathematics, and here, a neural network learns to do the same at the level of gestures and sounds.