Multiplexed subspaces route neural activity across brain-wide networks

One of the most fascinating things about the brain is its flexibility — how effortlessly it can switch between behaviors, thoughts, and goals. In our new paper, published in Nature Communications, we set out to understand how this happens at the level of neural networks. Using a combination of large-scale electrophysiology and cortex-wide calcium imaging in mice, we found that the brain routes information through what we call “multiplexed subspaces” — overlapping but distinct communication channels that link different regions. Each brain area participates in several of these networks at once, and which network is active can change from moment to moment, depending on what the brain needs to do.

What’s especially exciting is that this routing doesn’t require the brain to rewire itself. Instead, it seems to work geometrically: when neural activity within a region aligns with a particular “subspace,” that region becomes functionally connected with a specific network of other regions. By shifting the geometry of neural activity — rotating its representation within the brain’s high-dimensional activity space — the brain can instantly change which network it engages. It’s a remarkably efficient way to flexibly reconfigure computations on the fly.

We think this offers a new way to understand cognitive control — not just as a top-down signal, but as a dynamic reshaping of neural geometry that determines how information flows through the brain. It also suggests exciting parallels for artificial intelligence: perhaps the key to flexible, general-purpose AI isn’t adding more layers or parameters, but finding ways to dynamically reconfigure the geometry of internal representations, just like the brain does.