In computational neuroscience, connectomics, and neuromorphic engineering, researchers frequently rely on the static connectivity assumption: the simplification that the brain’s wiring diagram can be treated as a fixed parameter matrix. Whether mapping a structural connectome via electron microscopy or simulating thousands of spiking neurons, connections are traditionally treated as frozen snapshots.
However, biological brains are fundamentally non-stationary. Synapses continuously strengthen, weaken, form, and dissolve. Treating these connections as static parameters severely limits a model's long-term predictive power and prevents it from capturing how neural circuits adapt over time. To bridge this gap, a growing body of research focuses on integrating dynamic synaptic and structural plasticity directly into network inference, simulation, and optimization frameworks.
The Failure of Static Snapshots
When a model assumes that synaptic weights remain constant, it loses the ability to simulate realistic biological behavior over extended periods. For example, Marder (1993) demonstrated that biological circuits are continuously reconfigured by a chemical soup of neuromodulators. This allows a single physical circuit to switch between entirely different functional modes—a flexibility that traditional models with fixed connection weights fail to capture.
Even when we possess a complete, synaptic-resolution wiring diagram of an organism, a static map is insufficient. As Winding (2022) observed in the mapping of the Drosophila larval brain, fixed connectomes cannot fully predict functional states because they omit the real-time dynamics of neuromodulation and synaptic plasticity. Similarly, when building data-driven, whole-organism simulations, treating synaptic weights as static prevents the model from capturing the dynamic learning processes observed in living organisms over time, a limitation highlighted by Zhao et al. (2024) and Shlizerman et al. (2017).
The practical consequence of this omission is a decay in predictive accuracy. The authors of A framework for studying synaptic plasticity with neural spike train data (2014) showed that incorporating synaptic plasticity explicitly into statistical neural models significantly enhances their long-term predictive power. Static models inevitably fail as the underlying biological circuit reorganizes during an experiment.
Tracking Plasticity from Spiking Activity
To replace the static weight assumption with a dynamic one, we must be able to infer how connections change in real time using only observable data, such as neural spike trains. Traditional connectivity inference frameworks assume temporal stationarity, meaning they assume connection strengths do not fluctuate during a recording session. Toyoshima et al. (2024) noted that this assumption is frequently violated due to synaptic plasticity and neuromodulation.
To resolve this, Lemke (2022) developed a framework that tracks the temporal evolution of synaptic weights by combining long-term communication delays with short-term cross-covariance measures. This allows researchers to observe functional plasticity in action directly from spiking data.
Static Inference: [Spike Trains] ───> [Fixed Weight Matrix W]
Dynamic Inference: [Spike Trains] ───> [Time-Varying Weights W(t)]
When dealing with short, noisy recordings where spike counts exhibit high variance, standard Poisson models break down. She et al. (2021) addressed this by introducing an Empirical Bayes framework that models over-dispersion, successfully recovering time-varying functional weights where static models fail. On a macroscopic scale, Yang et al. (2025) demonstrated that multifractal analysis of a single neuron's interspike intervals can act as a "fingerprint" to infer hidden network architectures, though they emphasize that tracking how these signatures evolve during learning remains an open challenge.
Simulating Functional and Structural Plasticity
In forward simulations of neural circuits—whether in software or hardware—plasticity must be modeled on two distinct levels: functional plasticity (modifying the strength of existing connections) and structural plasticity (the physical creation and destruction of synapses).
In hardware emulation, such as FPGA-based motor nervous system simulators, Niu et al. (2012) pointed out that fixed connection weights prevent the system from simulating long-term biological learning or maladaptation following injury. In software simulations, relying on fixed, random internal weights ignores how active learning shapes network topology. Hiratani et al. (2012) showed that incorporating temporal plasticity mechanisms allows networks to naturally develop the heavy-tailed (lognormal) synaptic weight distributions observed in real brains, which inherently optimizes memory recall through stochastic resonance.
Structural plasticity takes this a step further by dynamically altering the network's topology. Instead of manually hard-wiring a static connectivity matrix, the authors of Automatic Generation of Connectivity for Large-Scale Neuronal Network Models through Structural Plasticity (2016) designed a system where networks grow and prune their own connections based on biological homeostatic rules. This is essential for realistic modeling because, as Parajuli et al. (2020) observed, dendritic spines undergo continuous morphological changes during learning. Static anatomical snapshots cannot capture how these geometric structures are dynamically recalibrated to maintain stable signal delivery.
Optimizing Plasticity Rules
If synaptic weights are treated as dynamic variables, what rules should govern their evolution? Rather than hard-coding rigid, hand-designed learning rules (like standard Hebbian learning), recent breakthroughs focus on optimizing the plasticity rules themselves using gradient descent.
Miconi et al. (2018) introduced differentiable plasticity, a framework where backpropagation is used to train the plasticity coefficients of individual connections. Instead of fixing the weights, the meta-optimization process determines how plastic a synapse should be based on experience. This allows the network to learn how to learn during its lifetime, significantly outperforming static-weight recurrent networks on memorization and few-shot learning tasks.
| Framework | Synaptic Weights | Plasticity Rules | Lifetime Adaptation |
|---|---|---|---|
| Traditional SNNs | Fixed after training | None | None |
| Fixed Plasticity | Dynamic | Fixed (e.g., standard STDP) | Rigid, non-optimal |
| Differentiable Plasticity | Dynamic | Optimized via Backpropagation | Context-aware, flexible |
To maximize data economy and energy efficiency, Rodriguez et al. (2022) introduced Short-Term Plasticity Neurons (STPNs). By training both the learning and forgetting rates per synapse via backpropagation, these networks manage their own short-term memory decay dynamically. This avoids the limitations of static decay parameters, allowing artificial networks to mimic the rapid, energy-efficient adaptation characteristic of biological brains.