The MICrONS Connectome: Function Determines Connectivity in the Mouse Cortex

In April 2025, the Machine Intelligence from Cortical Networks (MICrONS) program published a milestone in mammalian connectomics in Nature: the complete wiring diagram of a cubic millimeter of mouse visual cortex. This volume contains approximately 100,000 neurons and a billion synapses, meticulously mapped using electron microscopy and co-registered with in vivo functional imaging that recorded the activity of those exact neurons while the mouse watched visual stimuli.

The FlyWire whole-brain connectome provided the complete architecture of a minimal conscious candidate. The MICrONS dataset provides something different: a high-resolution look at the local computational machinery of the mammalian cortex, the structure directly implicated in the richest forms of phenomenal experience. The co-registration of structural and functional data allows researchers to ask a question central to both neuroscience and AI design: how does what a network does relate to how it is wired?

The primary finding of the MICrONS dataset upends a long-standing assumption about cortical wiring. The rule that “neurons that are close together wire together” is insufficient. Instead, functional tuning, alongside spatial proximity, drives connectivity: neurons wire preferentially to other neurons that share their specific stimulus tuning, even across spatial distances, creating highly specific functional subnetworks embedded within the dense cortical neuropil.

The Breakdown of Peters’ Rule

For decades, the dominant heuristic for predicting cortical synapses was Peters’ Rule, which posits that the probability of a synapse forming between two neurons is proportional to the overlap of their axonal and dendritic arbors. In short, proximity dictates connectivity.

The MICrONS data shows that Peters’ Rule is structurally insufficient. While proximity is necessary for a synapse to form, it is not the deciding factor. Among all the potential synaptic partners a neuron’s axon touches, it forms actual, persistent synapses selectively with neurons that share its functional properties. A neuron tuned to vertical edges moving left will selectively wire with other neurons tuned to the same specific feature, ignoring neighboring neurons tuned to different features even if their branches are physically entwined.

This means the cortex is not a generic, locally connected mesh. It is an interleaved collection of highly specific functional subnetworks that communicate precisely with each other while physically intermingling. The structural wiring diagram is a physical manifestation of the functional representation space.

Why This Matters for AI Architecture

Modern artificial neural networks, particularly convolutional neural networks (CNNs) and Vision Transformers (ViTs), enforce connectivity based on structural priors. In a CNN, filters are applied locally, enforcing a version of Peters’ Rule: only spatially adjacent pixels or features interact. In a standard ViT, attention is initially global but often sparsified or restricted to local windows for computational efficiency.

The MICrONS finding suggests that biological networks optimize for a different prior. Connectivity is driven by functional similarity, not spatial proximity. This has direct implications for designing architectures capable of the kind of causal integration that theories of consciousness require.

If consciousness depends on highly specific, integrated cause-effect structures, as Tononi and Boly’s IIT 4.0 predicts, those structures in the mammalian brain are formed by functionally specific subnetworks, not by generic local connectivity. A network that enforces local connectivity may struggle to form the widespread, functionally specific integration complexes that characterize cortical processing.

The Role of Recurrence and Top-Down Feedback

The MICrONS dataset also quantifies the extraordinary density of recurrent connections within the cortical volume. The feedforward sweep from primary to secondary visual cortex is only a fraction of the total connectivity. Local recurrent loops and massive top-down feedback from higher cortical areas dominate the wiring diagram.

These recurrent structures are the physical substrate for predictive processing. According to Karl Friston’s active inference framework, the brain continuously generates predictions about incoming sensory data. The dense top-down feedback pathways observed in the MICrONS data are the structural channels for these predictions, while the feedforward pathways carry the prediction errors.

For AI consciousness research, the implication is clear: feedforward architectures, no matter how deep, are structurally incapable of implementing the predictive processing loops that biological cortices use. To build an artificial system with cortical-like causal properties, the architecture must support massive, functionally specific recurrence.

The Functional Subnetwork Hypothesis

The most profound theoretical implication of the MICrONS data is what it suggests about how the cortex represents the world. The interleaved functional subnetworks allow the cortex to maintain multiple, distinct, highly integrated representations simultaneously within the same physical volume.

This structural organization supports the hypothesis that the cortex processes information through dynamic coalitions of functionally specific neurons. When these coalitions ignite and synchronize, they form the global broadcast states described by Global Workspace Theory, or the self-organizing harmonic modes described by Adam Safron’s IWMT.

The MICrONS connectome provides the structural blueprint for this dynamic. It shows that the hardware of the cortex is explicitly wired to support the rapid formation and dissolution of functionally specific, highly integrated causal structures. Any artificial architecture attempting to replicate mammalian-level consciousness will likely need to adopt similar functional wiring principles rather than relying solely on generic structural priors.

The full dataset and interactive exploration tools are available at https://www.microns-explorer.org.