A Connectome Is Not a Brain

Google Research, working with HHMI Janelia and other collaborators, has released a complete wiring diagram of the male fruit fly’s brain and central nervous system. I wrote about the fruit-fly connectome earlier, when I was more interested in the visual beauty of the map than in what it might actually teach us. The map contains more than 166,000 neurons and roughly 125 million synaptic connections. Researchers are already treating the connectome as a neural-network model: loading the wiring into simulations, tracing circuits, and asking what kinds of behaviour might emerge. This is fun, but a wiring diagram does not include everything that makes a biological brain work. Synaptic dynamics, neuromodulation, learning history, internal state and the body receiving the signals all matter. The connectome gives us a remarkably detailed structure, but structure is not yet a running brain.

That caveat takes me back to the argument around AI architecture. Yann LeCun has long argued that scaling today’s autoregressive architectures will not, by itself, get us to human-level intelligence, and has proposed a different path based on world models and planning. Geoffrey Hinton, by contrast, has often treated artificial neural networks as useful, if very rough, abstractions of biological brains. I am not trying to settle that argument here. I am also not sure that the choice is as clean as either side sometimes makes it sound. But the fruit-fly project is a useful reminder that even a real biological wiring diagram is not a complete explanation of intelligence. We should be careful when moving from “this resembles a brain” to “this will therefore produce the capabilities of a brain.”

A recent Nature Neuroscience paper makes the same point from another direction. Researchers recorded activity in the visual cortex of macaque monkeys while cameras tracked their spontaneous face and body movements. In the mouse studies that motivated the work, movement was associated with substantial activity in visual cortex. In the macaques, the apparent movement-related activity in V1, V2 and V3/V3A was largely explained by changes in the retinal image, especially when the eyes moved. Body movement itself contributed very little. The authors suggest that differences in anatomy, neuromodulation, behaviour and eye movements may help explain the contrast. For AI, the lesson is modest: even closely related biological systems can wire sensory and motor information together differently. A connectome may tell us what is connected, but the body, sensors and environment help explain why. That is the more interesting direction for AI research: design architectures together with the worlds they have to perceive and act in, rather than treating a network diagram as a universal blueprint.