Two birds with the same song in their heads. One sings while the other listens, then they swap.
Pitch against time, two to five kilohertz. Each bird’s notes are in its own colour.
One of each bird’s expectations about the song, as it runs. When the lines lie on top of each other, the two birds are in the same place in the same song.
Each bird’s expectations plotted against the other’s. On the diagonal they are identical. Apart, the points wander. Within earshot, they fall onto the line.
Friston and Frith give each bird the same generative model of a song: a slow attractor that shapes a fast one, which drives the voice. A bird either sings, acting so that its own predictions come true, or listens. It cannot do both well at once, so while it sings it turns down the precision of what it hears.
Because the two birds carry the same model, hearing each other is enough to bring their inner states into step. Communication, on this account, is that synchrony. Neither bird is copying the other. They are continuing one song between them.
Friston, K. & Frith, C. (2015). A Duet for one. Consciousness and Cognition. doi:10.1016/j.concog.2014.12.003. Open access, CC BY. Every quotation above is taken from the published text.
From the paper: two Lorenz attractors, the slow one about ten times slower, setting the manifold of the fast one; two fast states driving pitch (2 to 5 kHz) and loudness; turns of two seconds; auditory log precision 2 when listening and 0 when attenuated.
Simplified by me: the paper uses generalised predictive coding in SPM. This page uses a first-order version: each bird’s expectations follow its own model and are corrected by precision-weighted prediction error on what it hears, with the slow level explaining what the fast level cannot. The exact constants in the paper’s figure were not available to me as text, so the attractor settings are mine.
Checked before publishing (five random starts, sixteen seconds each). Within earshot the fast expectations synchronised in 5 of 5. Out of earshot they stayed apart in 4 of 5; in the fifth both birds happened to fall silent together. Without attenuation, sensation pulled on the singer 3.8 times harder. My first version failed the first check in 5 of 5 and was corrected. The slow level synchronises only partly here; in the paper it is identical.