For decades, the dominant metaphor for the brain has been the "wiring diagram." In this view, neurons are nodes and synapses are fixed edges, forming a stable architecture that dictates function. Most computational models—from the Blue Brain Project’s reconstructions to standard Spiking Neural Networks (SNNs)—rely on this static assumption: once the weights are set or the connections are mapped, the topology remains constant during the observation period.
However, biological reality is far more fluid. Synapses are not static copper wires; they are dynamic junctions that strengthen, weaken, appear, and vanish through synaptic and structural plasticity. A growing body of research argues that treating connectivity as a fixed parameter is a fundamental bottleneck. Instead, we must treat the network's structure as a continuous dynamical variable that evolves alongside the neural activity it generates.
The Failure of the Static Snapshot
Traditional "functional connectomics" often relies on a single point-in-time mapping. As Paul G. Fahey and colleagues note in Functional connectomics reveals general wiring rule in mouse visual cortex, these snapshots fail to account for how wiring rules reorganize during learning or in response to sensory changes.
When we assume a network is static, our inference algorithms often hallucinate connections or miss the "engine" of the system. For instance, Daniel Soudry points out in A shotgun sampling solution for the common input problem in neural connectivity inference that if connections evolve while we are observing them, our statistical reconstructions will be fundamentally skewed. We aren't just dealing with noise; we are dealing with a non-stationary target.
Inferring Hidden Plasticity from Activity
How do we "see" a connection change if we can only record spikes? The intuition is to move from correlation to prediction.
Kijung Yoon proposes a self-supervised Graph Neural Network (GNN) framework in Self-Supervised Discovery of Neural Circuits in Spatially Patterned Neural Responses with Graph Neural Networks that learns to predict future spikes. By forcing the model to explain the dynamics of the activity, the GNN uncovers the underlying synaptic structure even when many neurons are unobserved.
A more direct approach to capturing change is found in NetFormer, developed by Ziyu Lu and Hao Wang. In NetFormer: An interpretable model for recovering dynamical connectivity in neuronal population dynamics, they replace the standard "softmax" attention of transformers with a linearized mechanism. This allows the model to treat connectivity as a time-varying state. Instead of a single weight matrix, the model recovers a sequence of matrices, effectively "filming" the synaptic plasticity as it happens during a recording session.
Structural Plasticity as a Generative Rule
Beyond just inferring weights, some researchers are looking at the birth and death of the connections themselves—structural plasticity.
In Automatic Generation of Connectivity for Large-Scale Neuronal Network Models through Structural Plasticity, Sandra Diaz-Pier argues that we shouldn't "hard-wire" large-scale simulations. Instead, we should define biological growth rules that allow the network to prune and form synapses homeostatically. This shifts the focus from what the connections are to the rules that govern their change.
This perspective is echoed in the work of Minija Tamosiunaite and Florentin Wörgötter, who explore Temporally changing synaptic plasticity. They suggest that even the "learning rules" (like Hebbian plasticity) are not fixed; they evolve as a neuron matures.
Topological and Multifractal Signatures
If the network is constantly rewiring, how does it maintain any stable function? Researchers are finding answers in high-dimensional geometry and chaos theory.
- Topological Cliques: Pawel Dlotko and Henry Markram used algebraic topology to show that neurons organize into dense "cliques." In Topological analysis of the connectome of digital reconstructions of neural microcircuits, they suggest these structures are the true functional units, though they acknowledge that we still don't fully understand how these high-dimensional complexes adapt during learning.
- Multifractal Fingerprints: Ruochen Yang and Roozbeh Kiani demonstrated in Spiking dynamics of individual neurons reflect changes in the structure and function of neuronal networks that the timing of a single neuron's spikes contains a "fingerprint" of the surrounding network's topology. This implies that even if we can't see the whole network, the complexity of the spike train reveals the hidden structural state.
Toward Lifelong Learning Systems
The ultimate goal of this research direction is to move away from the "train-then-freeze" paradigm of artificial intelligence. By incorporating dynamic connectivity, we move toward systems capable of "lifelong learning."
As Feifei Zhao and Yi Zeng demonstrate with EB-NAS, brain-inspired architectures that evolve their own connectivity motifs are not only more accurate but more energy-efficient. The "mind control" framework proposed by John D. Medaglia in Mind control as a guide for the mind further suggests that if we can model how these "edges" evolve, we can develop better clinical interventions for neurological disorders by steering the network's structural evolution back toward a healthy state.