Head direction cells fire when an animal faces a particular direction. Point north: one set of neurons fires. Point south: a different set. The cells function as an internal compass, and decades of laboratory research have characterized their tuning curves, their interactions with place cells, and their role in spatial navigation. The laboratory version of this system is clean and well-understood. The animal moves through a small, familiar space. The cells fire reliably. The compass works.
Palgi, Ulanovsky, and colleagues (Science, October 2025) recorded head direction cells from fruit bats flying freely on Latham Island, a remote oceanic island 25 miles off Tanzania. The team implanted wireless neural recording devices — the smallest of their kind — in six bats and tracked both brain activity and GPS position as the animals navigated the island over multiple nights. The results differed from laboratory recordings in ways that weren't predicted by the laboratory model.
On the first nights, head direction cells fired imprecisely. By night five or six, as the bats learned the island's landmarks — coastline, perches, the researchers' tents — the cells stabilized to fire in coordination with precise directions. The compass calibrated itself through exploration. In the lab, animals are already familiar with their environment. You never see the calibration process because the controlled conditions eliminate it. The stabilization is invisible not because it doesn't exist, but because the lab removes its prerequisite: an unfamiliar, large-scale space that takes days to learn.
Preliminary data presented at the Society for Neuroscience meeting revealed a further discrepancy. In the wild, place cells didn't just record location — they also encoded how fast the bat was flying. Speed information was superimposed on spatial information in the same cells. This multiplexing doesn't appear in laboratory recordings because lab animals aren't flying at natural speeds through kilometer-scale environments. The computation exists, but the controlled environment doesn't provide the inputs that produce it.
The standard defense of laboratory neuroscience is that controlled conditions isolate the signal from noise. Remove extraneous variables, and what remains is the mechanism in its pure form. The bat data invert this assumption. The “extraneous variables” — unfamiliar terrain, natural flight speeds, large-scale landmarks — aren't noise around the signal. They are the inputs that produce computations the signal includes. Remove the variables, and you don't get a cleaner signal. You get a different, simpler computation that omits what the system actually does.
The general pattern: when a system's computation depends on its inputs, simplifying the inputs doesn't isolate the computation — it replaces it with a simpler one. The lab bat and the wild bat use the same neurons. The difference is what those neurons are asked to do. In a small box, a head direction cell tracks direction. On an island, it tracks direction while calibrating to novel landmarks over days and encoding flight speed. The controlled environment didn't remove noise from the compass. It removed the conditions under which the compass learns, stabilizes, and multiplexes. The simplification is the perturbation.