BK channels are potassium channels distinguished by their exceptionally high unitary conductance, meaning the rate at which potassium ions can pass through an individual channel molecule. Across the animal kingdom, the source release reports that their conductance can be roughly 10 to 20 times that of other types of potassium channels, depending on the comparison. Their unusual conductive capacity has long been known and can now be explained in structural terms, but the adaptive reason for the trait remained unclear. The new study addresses that question by connecting three features: the evolution of conductance across species, the calcium requirements for opening the channels and the precision of neuronal spike timing.

To identify the evolutionary pattern, OIST evolutionary biologists Thomas Bourguignon and Simon Hellemans worked with computational neuroscientist Erik De Schutter and colleagues in the Evolutionary Genomics and Computational Neuroscience units. The team used genomic information and published measurements of BK-channel conductance to map channel properties onto the evolutionary relationships among animals. This comparison showed that conductance is not uniform across species. Mammals and birds were at the high-conductance end of the range, while other vertebrates and invertebrates generally showed lower values. Mollusks and insects were among the groups represented at some of the lowest conductance levels. The analysis also found a broad trend in which species shown as closer to mammals on the evolutionary tree tended to have higher BK conductance.

The cross-species mapping established the pattern that required a functional explanation, but it did not by itself reveal why higher conductance might be advantageous. The researchers therefore turned to STEPS, a molecular simulation program that models ion channels and their surrounding cellular environment. They represented individual calcium-activated BK channels with different conductance levels and included the effects of intracellular calcium concentration. BK channels require four bound calcium ions to open fully, an important detail because each activated channel binds calcium and thereby reduces the amount of free calcium remaining to activate other channels.

In the molecular simulations, calcium removal emerged as a key limitation on collective BK-channel activation. At a low unitary conductance, many more channels had to open to produce the same overall rate of potassium flow through the membrane. As calcium ions became bound to those channels, the available free calcium declined, so the channels effectively competed with one another for activation. The model did not predict the same degree of limitation at higher unitary conductance, as found in mammals and birds. At that higher conductance, fewer open channels were needed to achieve the same overall conductance. The simulations therefore suggest that greater current per open molecule can make BK-channel activation more efficient by easing competition for free calcium.

That result is a mechanistic estimate from the molecular models rather than a direct measurement of channel behavior across the species examined. It also depends on the calcium concentrations and other assumptions used in the simulation. Even so, the model provides a physical link between the observed evolutionary pattern and a cellular constraint. The important distinction is not simply that a high-conductance channel transports more potassium. It is that a channel carrying more current per molecule can contribute the required total current without requiring the simultaneous binding of calcium by as many channel molecules. In this way, the model connects a microscopic transport property with the limited availability of an intracellular activator.

The team next used a second computational model to test the functional consequences of these conductance differences for spike timing, the precise moments when neurons produce electrical signals. This neuronal model focused on a Purkinje cell, a type of neuron in the cerebellum, and represented the effects of lower and higher BK-channel conductance. Under the modeled conditions, lower unitary conductance was associated with greater variability in the timing of neuronal spikes, while higher conductance supported more precise timing. The result offers a possible explanation for why selection in more complex nervous systems might favor BK channels with higher conductance: the same cellular mechanism can reduce activation limitations while producing less variable timing in the model.

Precise spike timing is relevant to the temporal coordination of electrical signals that supports functions including motor control, memory retention and sensory processing. The study’s authors therefore suggest that increasingly complex motor functions may have made precise, sub-millisecond spike timing more important during the evolution of mammals and birds. If improved timing contributed to neural performance, variants with higher-conductance BK channels could have had an advantage. This remains an evolutionary hypothesis tested through a combined computational framework. The models do not directly show that ancestral animals altered their channels in response to behavioral demands, and they do not reproduce every neuronal type, circuit property or physiological condition found in living animals.

Several limitations define the scope of the findings. The neuronal simulations are theoretical approximations built around a single cell type, the Purkinje cell, rather than a complete nervous system. The molecular simulations likewise depend on selected parameters, including calcium concentrations, to represent channel behavior. The evolutionary comparison also combines species with distinct evolutionary histories, so the observed association between conductance and relatedness does not isolate adaptation from every other influence on the trait. BK channels are also widespread across cell types and participate in functions beyond neuronal spike timing, including muscle contraction. The study does not establish a medical application, a clinical benefit or a universal model of potassium-channel evolution. Its strongest conclusion is that its models provide a coherent and testable explanation for why unusually high BK conductance may have evolved.

Despite these limitations, the work links levels of biological organization that are often studied separately. Comparative analysis identifies the pattern of higher conductance in mammals and birds, molecular simulation estimates how conductance changes competition for calcium, and neuronal modeling tests the resulting effect on spike timing. The approach turns a longstanding question into a quantitative proposal: greater BK-channel conductance may have evolved because it permits the same total potassium current with fewer calcium-bound channels, helping neuronal firing remain precise as the functions supported by nervous systems became more demanding. The result does not close every question about the channels, particularly their roles in other cell types and biological processes, but it offers a concrete mechanism for a puzzle that structural explanations alone could not resolve.

Subject of Research: Evolutionary and computational analysis of BK potassium-channel conductance and neuronal spike timing

Article Title: More current with less calcium: Computational models help solve longstanding ion channel evolution mystery

Article References: More current with less calcium: Computational models help solve longstanding ion channel evolution mystery. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: BK channels, ion conductance, potassium channels, neuronal timing, calcium activation, molecular simulation, evolutionary genomics, computational neuroscience, Purkinje cells, animal evolution, current, less

The research team, led by Bartłomiej Król-Józaga of AGH University of Krakow together with colleagues at Dartmouth College and Jagiellonian University, tackled a long-standing gap in computational neuroscience. Most existing models of decision making simulate only one of the many parallel operations the brain performs—reward processing, evidence accumulation, strategy selection, or option choice—rather than the whole coordinated process. An earlier spiking neural network model had attempted a more holistic simulation, but it ignored a key biological feature: the powerful role that inhibitory neural connections play in shaping cognition.

Inhibition is far more than a neural brake. Inhibitory interneurons synchronize the brain oscillations that coordinate the networks involved in choice, filter out irrelevant information, and refine neural responses so that behavior can adapt to varying cognitive demands. This mechanism supports the interplay between brain regions central to decision making: the orbitofrontal cortex (OFC), which binds sensory input with value signals; the dorsolateral prefrontal cortex (dlPFC), which accumulates and integrates information over time; and the basal ganglia, which select the best option. Without inhibition, neural circuits risk runaway excitation—the kind of pathological over-activation seen in conditions such as epilepsy and schizophrenia.

To test whether inhibition could explain how people choose decision strategies, the team used a well-studied experimental paradigm called the sequential probabilistic inference task. Twenty-four participants, aged 21 to 27, memorized the importance weights of six attributes and then viewed the binary values of two choice options, one attribute at a time. On each trial they could either pick option A, pick option B, or request more information. Crucially, the task came in two flavors: a Non-Compensatory condition, where one attribute dominates and a simple heuristic called Take The Best (TTB) works well, and a Compensatory condition, where all attributes carry similar weight and the rational Weighted Additive rule (WADD)—which sums weighted evidence across everything—is the smarter strategy.

The model itself was built within the Neural Engineering Framework (NEF), a mathematical methodology for constructing large-scale neural models with realistic spiking neurons. In the NEF, each neuron is characterized by an encoder, a gain, and a bias that together define its tuning curve—how strongly it fires in response to an input signal. Filtered spike trains are decoded with optimized weights to reconstruct the represented signals, and transformations such as value-weight multiplication are implemented through learned connection strengths. Dynamics, the third NEF principle, captures how neural systems evolve over time, enabling the model to reproduce mechanisms like working memory, evidence accumulation, and choice competition.

The architecture mirrors known brain anatomy. Cue values flow through perception and memory components into the model’s OFC population, where they are combined with stored importance weights. The weighted values then feed into a dlPFC evidence population whose recurrent connections act as an integrator, accumulating information across time steps. A dlPFC contrast population computes the difference between the integrated values of the two options, sharpening the distinction before signals reach the basal ganglia. There, a Winner-Take-All mechanism—implemented through the striatum, globus pallidus, and subthalamic nucleus—selects the option with the highest accumulated evidence, triggering a motor response once the decision threshold is reached.

The critical innovation was a neuromodulatory pathway inspired by the locus coeruleus-norepinephrine (LC-NE) system, a brainstem nucleus famous for regulating attention, arousal, and cognitive effort. In the model, a neuromodulatory population receives input from the accumulating dlPFC evidence—interpreted as a marker of cognitive load—and sends inhibitory connections to the OFC. As accumulated evidence grows, this pathway progressively suppresses the incoming cue-value and weight products, simulating how effort-related arousal sharpens the selectivity of pre-decisional information processing. The strength of this inhibition is captured by a single participant-specific parameter, fitted to each individual’s behavioral data alongside parameters for decision threshold, urgency scaling, and lateral-inhibition gain.

The validation strategy was deliberately demanding. The researchers fitted both model variants—the standard version and the inhibition-enhanced version—only to the Non-Compensatory condition, then transferred the frozen parameters to the structurally different Compensatory condition as an out-of-distribution test. This design served a dual purpose: it directly probed whether the load-sensitive inhibitory mechanism, calibrated under low cognitive load, could predict behavior under higher load, and it acted as a built-in control for model complexity. A model that merely exploited an extra free parameter to overfit the training data should generalize worse, not better, to unseen conditions.

The results were striking. On the training condition, the two models performed comparably, with no statistically significant difference. But on the held-out Compensatory data, the inhibition model achieved approximately 17 percent lower mean prediction error than the no-inhibition model—0.55 versus 0.67—a difference confirmed by a Wilcoxon signed-rank test (p = 0.009, rank-biserial correlation r = 0.55), with the inhibition model winning in 16 of 24 participants. Because the advantage appeared precisely where added flexibility should have hurt rather than helped, the finding argues strongly against a mere complexity artifact. Parameter importance analysis reinforced this picture: the inhibition strength ranked second on average, but with high inter-individual variability, being decisive for some participants and negligible for others—hinting that arousal-based inhibition shapes decision behavior meaningfully in only a subset of people.

Intriguingly, the inhibition advantage showed up in the continuous pattern of information sampling rather than in categorical strategy shifts. Both model variants reproduced the characteristic bimodal shape of human cue acquisition—people tend to either sample very few cues or exhaust all six—and both predicted similar WADD-versus-TTB strategy distributions across conditions. What inhibition bought was a quantitatively more accurate replication of how many attributes people consulted under demanding conditions, consistent with the idea that neuromodulation coupled with inhibition grades the selectivity of information gathering rather than flipping the brain between discrete strategies. The authors note that future work should refine the cognitive-load signal to produce realistic trial-by-trial buildup, validate neural trajectories against pupillometry and fMRI data, and explore the role of other neuromodulators such as dopamine. For now, the study offers compelling evidence that the brain’s inhibitory circuitry is not just noise suppression—it is an active, load-sensitive mechanism that helps us decide how to decide.

Subject of Research: The role of neural inhibition in decision strategy selection modeled with spiking neural networks

Article Title: The role of inhibition in modeling decision making with spiking neural networks

Article References: Król-Józaga, B., Duggins, P., Broniec-Wójcik, A., & Wichary, S. (2026). The role of inhibition in modeling decision making with spiking neural networks. Frontiers in Neuroinformatics, 20, Article 1854811. https://doi.org/10.3389/fninf.2026.1854811

Image Credits: AI Generated

DOI: 10.3389/fninf.2026.1854811

Keywords: spiking neural networks, decision making, neural inhibition, orbitofrontal cortex, locus coeruleus, norepinephrine, cognitive load, heuristics, Neural Engineering Framework, basal ganglia, evidence accumulation, computational neuroscience

The editorial, written by José Roberto Castilho Piqueira and Antonio C. Roque of the University of São Paulo and Renato Teodoro Ramos of the University of Toronto, frames the challenge with unusual clarity. Cognition, they argue, depends on processes occurring across multiple spatial and temporal scales, and theoretical and computational approaches are essential not merely for describing neural activity but for identifying principles that connect these levels of organization and ultimately relate neural dynamics to perception, memory, and behavior. The five studies they highlight range from statistical frameworks for massive electrophysiological recordings to mathematical reviews of differential-equation models and sweeping conceptual proposals about neural communication and emergence. Despite their methodological diversity, all five converge on a single question: how can patterns of neural activity and interaction across scales give rise to organized brain function?

Two of the contributions attack this question head-on using large-scale neural recordings. The first, by Bong and colleagues, addresses a problem that has grown acute as recording technology has advanced: modern experiments can capture oscillatory activity from many brain regions simultaneously, but conventional analysis often reduces these multichannel recordings to single regional signals, discarding the rich structure of interactions between populations. The researchers developed a statistical framework for identifying time-varying relationships between oscillatory activity recorded simultaneously from different brain regions. Rather than collapsing the data, their approach extracts latent components whose amplitude fluctuations reveal inter-regional interactions and their lead-lag structure — that is, which population’s activity tends to precede the other’s in time.

The power of this framework became apparent when the team applied it to local field potentials recorded from prefrontal cortex and visual area V4 during a memory task. The method provided a way of characterizing dynamic communication between high-dimensional neuronal populations, and, crucially, it could identify amplitude-based relationships even in the absence of phase locking or coherence. This is a significant technical advance. Many conventional measures of oscillatory coordination rely on the assumption that interacting brain regions maintain a stable phase relationship, but real neural communication often violates that assumption. By detecting amplitude couplings that other methods miss, the approach complements rather than replaces existing tools, offering a more complete picture of how prefrontal and visual areas talk to each other during working memory.

The second recording-based study, by Olarinre and colleagues, approaches inter-regional coordination from a different angle: transient population events rather than continuous oscillations. Using large-scale Neuropixels recordings — dense silicon probes that can monitor hundreds of neurons across multiple areas at once — from several visual cortical areas and the thalamus, the researchers built a computationally efficient framework for estimating, on individual trials, the timing of stimulus-evoked population bursts and the correlations between those timings across regions. This trial-by-trial perspective matters because averaged activity can hide precisely the variability that carries information about how brain regions coordinate.

Their results revealed both regularities consistent with the known organization of the visual system and substantial animal-to-animal variability. In particular, correlations among cortical areas tended to be stronger than correlations between cortex and thalamus, a finding that aligns with the hierarchical structure of visual processing pathways. Together, the first two studies drive home a point that is reshaping the field: cognition-related neural coordination cannot be understood solely from average activity. Temporal variability, population structure, and dynamically changing relationships between brain regions contain important information that traditional averaging approaches discard.

A third contribution takes a deliberately speculative and integrative stance. Rather than beginning with a particular dataset, Baker and Cariani develop a broad hypothesis about how information might be encoded, distributed, transformed, and communicated in neural systems. Their framework combines traveling waves, temporal spike patterns, phase relationships, synchronization, recurrent networks, and distributed representations with concepts borrowed from wave physics, holography, and communication theory — including the intriguing notion of single sideband communications, a technique from radio engineering that could, in principle, allow neural systems to carry temporally structured information efficiently. In this view, oscillations and traveling waves are not merely correlates of cognition, the way a shadow is a correlate of a body, but potential mechanisms for coordinating distributed neuronal populations and carrying information across the brain. The proposal is deliberately integrative and generates testable hypotheses that can be confronted with increasingly detailed measurements of spatiotemporal neural dynamics — measurements that next-generation recording technologies are just beginning to deliver.

The fourth study tackles a question that is central to all of computational neuroscience: how can descriptions at different scales be connected mathematically? Çetin and colleagues provide a review that organizes differential-equation-based approaches into three interconnected families. Ordinary differential equations provide mechanistic descriptions ranging from single-neuron models — such as the classic Hodgkin-Huxley equations that describe how ion currents generate action potentials — to neural-mass and whole-brain systems that treat entire regions as unified units. Stochastic formulations incorporate the variability that is intrinsic to neuronal activity, acknowledging that real neurons are noisy, not deterministic machines. Mean-field partial differential equations, finally, describe collective dynamics in large neuronal populations, capturing how the distribution of neurons across states evolves over time. Viewed together, these approaches provide mathematical bridges between microscopic neuronal mechanisms and mesoscopic or macroscopic phenomena such as oscillations, synchronization, multistability, and pattern formation. The authors also highlight a critical future challenge: integrating models across scales more systematically and connecting them more tightly with experimental data and modern methods of inference and machine learning.

The fifth contribution broadens the multiscale perspective in a way that challenges a deep assumption of mainstream neuroscience. Oruro and colleagues argue that neuronal morphology — the intricate branching shape of dendrites and axons — should be regarded as a dynamic level of organization in its own right, not simply as a fixed substrate on which neural computation occurs. Their framework connects neuronal form, circuit organization, behavior, and emergent function, drawing examples from development, mother-infant interactions, epilepsy, and neuropsychiatric conditions. This perspective highlights a fundamental feature of brain-based cognition: causal relationships across levels are potentially reciprocal. Cellular and morphological processes shape network dynamics and behavior, while higher-level interactions and experience can, in turn, influence lower-level organization. Experience-dependent changes in dendritic structure, for instance, feed back into the very network dynamics that generated the experience. Computational modeling, the authors contend, offers an important means of investigating these cross-scale relationships that run in both directions simultaneously.

Taken together, the five contributions deliver a message that is both humbling and energizing: there is unlikely to be a single privileged scale or methodology for understanding brain-based cognition. Statistical methods are needed to extract meaningful interactions from increasingly high-dimensional recordings; dynamical models provide mechanistic links between neuronal and population activity; concepts from physics and information processing suggest possible organizing principles for neural communication; and complex-systems approaches emphasize emergence and reciprocal interactions across levels of organization. The challenge ahead, the editors conclude, is to bring these perspectives into closer contact with one another and with experimental data. Progress toward an explanation of cognition will depend not only on increasingly detailed measurements or increasingly sophisticated models, but on the ability to connect structure, dynamics, information, and function across scales. The articles collected here provide complementary steps toward that broader theoretical and computational synthesis — and they suggest that the road to understanding the mind runs not through any single theory, but through the disciplined conversation between them.

Subject of Research: Editorial: Theoretical and computational insights into brain-based cognition

Article Title: Editorial: Theoretical and computational insights into brain-based cognition

Article References: Piqueira, J. R. C., Roque, A. C., & Ramos, R. T. (2026). Editorial: Theoretical and computational insights into brain-based cognition. Frontiers in Computational Neuroscience, 20, Article 1986598. https://doi.org/10.3389/fncom.2026.1986598

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1986598

Keywords: computational neuroscience, cognition, brain dynamics, neural oscillations, Neuropixels, traveling waves, mean-field models, neural communication, emergence, multiscale modeling, prefrontal cortex, visual cortex

The team, led by Claus Metzner, Achim Schilling, Andreas Maier, and Patrick Krauss, focused on three structural regularities that are hallmarks of biological networks. The first is Dale’s principle, the observation that a biological neuron is either excitatory or inhibitory, but never both, so that all of its outgoing connections share a single polarity. The second is reciprocity: if neuron A projects to neuron B, then B is more likely than chance to project back to A, a bias well documented in cortical circuits and central to Hopfield networks. The third is modularity, the tendency of the brain to organize neurons into densely connected groups embedded in a sparser background of weaker connectivity, a pattern that recurs from microcircuits up to whole brain areas.

Earlier work by the same group had shown that the dynamical regime of an RNN can be steered by purely statistical properties of its weight matrix: the width of the weight distribution, the connection density, and the balance between excitatory and inhibitory weights together determine whether the network settles into fixed points, oscillates globally, or descends into chaos. But those studies assumed statistically independent weights, an analytically convenient simplification that real brains violate. The new study therefore built weight matrices in which each of the three regularities could be dialed up continuously through its own control parameter, allowing the researchers to isolate the effect of each regularity while holding the fundamental statistics constant.

To characterize the resulting dynamics, the team used three compact measures. The fluctuation measure captures the average temporal variability of neural activations, distinguishing quiescent networks from strongly oscillating ones. The covariance measure, deliberately defined without mean subtraction or normalization, distinguishes global oscillations, where successive activations flip sign, from fixed points, where activations stay constant, and from chaos, where the average product hovers near zero. The nonlinearity measure indicates whether neurons operate in the linear middle of their sigmoidal activation function or are driven into saturated, digital-like states. Validation scans confirmed that each control parameter reliably produced its intended structural feature without substantially distorting the others.

The baseline phase diagrams, mapped across the plane spanned by coupling strength and excitatory-inhibitory balance, revealed four familiar regimes: a quiescent region at weak coupling, an oscillatory wing dominated by strong inhibitory connections, a fixed-point wing dominated by strong excitatory connections, and a chaotic regime near the balanced center. When the network was then embedded in a reservoir computing framework and asked to generate predefined output sequences after being primed by class-specific inputs, accuracy stayed high in the quiescent regime, collapsed in the chaotic regime, and also degraded in the deeply saturated parts of the oscillatory and fixed-point regions. Only narrow bands at the edges of chaos retained near-perfect performance, and those bands shrank as coupling strength increased.

The surprises came when the regularities were switched on. Strong Hopfield reciprocity, which in classical Hopfield networks stabilizes stored patterns as fixed-point attractors, turned out to be harmful here. It shrank the quiescent regime, pushed neurons toward saturated, quasi-digital operation, and drove accuracy in the chaotic regime even lower than in the standard random network. Correlation analysis explained why: reciprocal coupling produced strongly correlated pairs and blocks of neurons, some locked into period-two oscillations and others into fixed points, with near-perfect Pearson correlations that persisted even under added noise. The authors note that their use of synchronous updates, rather than the sequential updates of the original Hopfield model, explains why reciprocity produced oscillatory rather than memory-like attractors.

Dale homogeneity, by contrast, proved beneficial. When every sending neuron was constrained to a single output polarity, the initially weak correlations of the chaotic state gave way to stronger, predominantly positive correlations, and the activations settled into quasi-periodic collective fluctuations with noticeable phase shifts between neurons. Crucially, task accuracy improved across all tested phase points, and the reservoir could tolerate significantly higher coupling strengths within the chaotic regime before performance broke down. The structured, directional signal flow that Dale’s principle imposes appears to tame erratic fluctuations without sacrificing the dynamical richness that reservoir computing needs.

The most dramatic effects belonged to modularity. When the weight matrix was partitioned into square blocks, some assigned strong weights and others weak ones while the overall distribution width was kept constant, the oscillatory and fixed-point regimes shrank to narrow slivers of phase space, and the chaotic regime was replaced by a broad region of moderate, irregular but non-repetitive dynamics. Accuracy then rose to near-maximal levels across almost the entire phase diagram, meaning that even previously unproductive dynamical regimes became computationally useful. The correlation matrices showed internally coherent neuron clusters with weaker, mixed-sign coupling between modules, a balanced structure the authors identify as particularly favorable for reservoir computing. The gain was confirmed in a larger 100-neuron reservoir and in an independent patch-classification task, where accuracy climbed monotonically from near chance level as modularity increased.

Interestingly, the benefits of modularity did not require true block structure. When the block size was reduced so that no distinct modules remained, increasing the modularity parameter merely transformed the weight distribution from a single Gaussian into a heavy-tailed mixture of two Gaussians, and this alone still improved accuracy at one phase point. The authors draw a parallel to the human cortex, where synaptic strengths follow approximately log-normal, heavy-tailed distributions, a heterogeneity thought to combine stable high-impact pathways with a flexible background of weak connections. Dale homogeneity and modularity thus echo design principles that evolution appears to have already discovered.

The study has clear boundaries: it examined a single benchmark task, used fixed activation functions and untrained recurrent weights, and varied each regularity in isolation. Whether Dale homogeneity and modularity act synergistically, or whether reciprocity counteracts their benefits, remains open, as does the question of how these effects scale to large, hierarchical architectures with plastic synapses. The authors also suggest applying the same dynamical metrics to empirically derived connectomes of organisms such as C. elegans, Drosophila, or zebrafish larvae, and propose moving beyond eigenvalue-based criteria toward measures of trajectory reproducibility and state-space geometry. If the findings generalize, the message is striking: the structural quirks of biological wiring are not idiosyncrasies to be smoothed away, but functional ingredients that artificial networks may be missing.

Subject of Research: Effects of Dale's principle, reciprocity, and modularity on the dynamics and reservoir computing performance of recurrent neural networks

Article Title: Organizational regularities in recurrent neural networks

Article References: Metzner, C., Schilling, A., Maier, A., & Krauss, P. (2026). Organizational regularities in recurrent neural networks. Frontiers in Complex Systems, 3, Article 1636222. https://doi.org/10.3389/fcpxs.2025.1636222

Image Credits: AI Generated

DOI: 10.3389/fcpxs.2025.1636222

Keywords: recurrent neural networks, reservoir computing, Dale's principle, Hopfield networks, modularity, neural dynamics, chaos, edge of chaos, connectomics, computational neuroscience, weight matrices, machine learning

The research, conducted at Albstadt-Sigmaringen University and Chemnitz University of Technology, tackles a tension that has long lurked beneath the surface of Hebbian convolutional networks. These models, which learn features through local synaptic plasticity rules inspired by the pioneering work of Donald Hebb, have shown impressive performance on visual classification benchmarks. But their success has always been coupled to a specific architectural trick: exact translational weight sharing, where the same filter is copied across every spatial position in an image. While this prior is computationally efficient and statistically powerful, it is not anatomically literal. Real neurons in the visual cortex do not literally share identical synaptic parameters across retinotopic positions, and the new study asks whether useful visual representations can still emerge when this crutch is removed.

To answer this question, the researchers turned to Hebbian principal component analysis, or HPCA, a local learning rule rooted in Oja’s classic formulation of synaptic plasticity. The rule combines a Hebbian potentiation term — synapses strengthen when presynaptic and postsynaptic activity coincide — with a projection-dependent stabilizing correction that prevents unbounded growth. Competition among hidden units is introduced through population-wise divisive normalization, a mechanism inspired by canonical normalization models in sensory neuroscience. The result is a learning rule that extracts principal components from input statistics using only information available at or near the synapse, a constraint that aligns far more closely with biological reality than the global error signals of backpropagation.

The experimental design was elegantly simple. The researchers compared three architectural regimes on CIFAR-10, MNIST, and STL-10. A fully connected network served as a non-spatial baseline. A position-specific locally connected architecture, developed specifically for this study, preserved spatially restricted receptive fields while assigning each unit its own unique synaptic weights — no sharing whatsoever. Finally, HPCA was substituted into two established shared-kernel convolutional Hebbian benchmarks, the modular Hebbian-CNN framework and the original SoftHebb CNN, to test whether the same learning principle remains effective when exact weight sharing is restored.

The results were striking. On CIFAR-10, the position-specific locally connected HPCA model significantly outperformed the fully connected reference, achieving 60.56 percent accuracy compared to 52.74 percent — a difference of 7.82 percentage points that held up under rigorous statistical testing with paired-samples t-tests across ten independently initialized runs. Even more remarkably, when the researchers disabled data augmentation in both conditions, the locally connected advantage persisted at 7.52 percentage points, confirming that the benefit was architectural rather than an artifact of transformation-rich training. A plateau-triggered optimizer-switch variant pushed the best locally connected result to 61.42 percent, placing it in the same range as earlier convolutional HPCA models that relied on shared kernels.

The shared-kernel experiments revealed a different story. When HPCA was embedded in the modular Hebbian-CNN benchmark, it reached 74.51 percent under common preprocessing and 76.57 percent in the SoftHebb architecture under reference preprocessing. These numbers substantially exceeded the locally connected results, confirming that exact weight sharing provides a powerful statistical prior for natural-image tasks. However, the researchers were careful to note that this gap cannot be attributed to weight sharing alone. Depth, hierarchy, preprocessing, competition mechanisms, learning-rate schedules, and readout protocols all co-varied across the configurations, making it impossible to isolate any single factor as the sole cause of the performance difference.

One of the most important contributions of this work is its systematic sensitivity analysis of the convolutional benchmarks themselves. The researchers discovered that preprocessing choices had dramatically different effects depending on the benchmark. In the modular Hebbian-CNN framework, switching to the common HPCA preprocessing pipeline increased accuracy by 4.01 percentage points. In the SoftHebb architecture, the same switch actually reduced accuracy by 0.54 percentage points. This protocol dependence was not specific to the HPCA substitution — even the reproduced SoftHebb reference conditions showed sensitivity to learning-rate and schedule changes, failing to recover the published reference accuracy. The findings suggest that raw accuracy differences between Hebbian models may reflect the entire benchmark configuration rather than the intrinsic quality of any particular learning rule.

To address concerns about parameter count, the researchers conducted a careful control experiment. They reduced the locally connected readout from 2,048 to 208 GELU units, matching the approximately 5.9 million parameter scale of the shared-kernel architectures. The parameter-matched locally connected model achieved 59.16 percent, only about 1.4 percentage points below the standard configuration. Yet the shared-kernel models still outperformed it by 15 to 17 percentage points under matched preprocessing. Because this advantage far exceeds the cost of parameter reduction, the researchers concluded that the performance gap does not arise from differences in total model capacity but from the structural benefits of weight sharing and architectural depth.

Receptive-field visualizations provided additional insight into what the different architectures actually learned. The locally connected HPCA model produced heterogeneous, position-specific filters consisting of small-scale chromatic mixtures, local contrast variations, and patch-like high-frequency structure — with no repeated template across the image, consistent with the absence of weight sharing. In contrast, the shared-kernel HPCA configurations exhibited a clear layer-wise progression from coarse block-like chromatic filters in the first layer to more diffuse color-texture combinations in the second and sparse, distributed high-frequency patterns in the third. This hierarchical organization mirrors the kind of feature composition that makes convolutional networks so effective, and it emerged spontaneously from local Hebbian learning when the shared-kernel prior was available.

The broader implications of this study extend beyond benchmark comparisons. By demonstrating that useful frozen visual representations can be acquired without exact translational weight sharing, the researchers have established a stricter test case for biologically constrained visual learning. The position-specific locally connected HPCA architecture represents a middle ground between fully connected networks, which ignore image structure entirely, and convolutional networks, which impose a strong shared-kernel prior that real neural circuits do not literally implement. The work also raises intriguing questions about how biological systems might achieve convolution-like behavior without explicit parameter tying — perhaps through transformation-rich visual experience, lateral Hebbian mechanisms that promote approximate weight alignment, or structural plasticity that adapts receptive-field topology during development. As the field continues to grapple with the gap between biologically plausible learning and state-of-the-art performance, this study provides both a valuable controlled model and a clear roadmap for future investigations into how local plasticity, spatial connectivity, and architectural inductive biases interact to shape visual representation learning in brains and machines alike.

Subject of Research: Hebbian principal-component learning in biologically constrained visual neural network architectures

Article Title: Locality and weight sharing shape Hebbian principal-component learning in biologically constrained visual models

Article References: Inoue, P., Röhrbein, F., & Knoblauch, A. (2026). Locality and weight sharing shape Hebbian principal-component learning in biologically constrained visual models. Frontiers in Computational Neuroscience, 20, Article 1943064. https://doi.org/10.3389/fncom.2026.1943064

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1943064

Keywords: Hebbian learning, local plasticity, weight sharing, visual representation learning, receptive fields, convolutional neural networks, biological constraints, principal component analysis, CIFAR-10, computational neuroscience, neural computation, unsupervised learning

Detailed neuronal models have become essential instruments in modern neuroscience. They allow researchers to examine the mechanisms behind observed electrical behavior and to predict how neurons respond under conditions that are difficult or impossible to study directly in experiments. They also serve as the fundamental building blocks of large-scale network models of brain regions. Yet many earlier models suffered from significant limitations: they were often fitted to a narrow set of experimental conditions, leaving little information about how they would behave outside the scope of the original study. Because multiple parameter combinations can fit limited data equally well, such underconstrained models can produce unrealistic parameter values and incorrect predictions when applied to new situations.

The motivation for the new work came in part from a sobering assessment published a few years ago. Using the HippoUnit validation tool, researchers compared several available rat CA1 pyramidal cell models against experimental data from a variety of paradigms and found that although some models captured certain aspects of the cells’ behavior—typically those they were originally built to replicate—none could simultaneously reproduce all the relevant behaviors. A thoroughly validated, accurate, general-purpose model of this crucial cell type was simply missing. The new study set out to fill that gap by leveraging the extraordinary wealth of anatomical, biophysical, and physiological data available for CA1 pyramidal cells.

The construction process was systematic and iterative. The team started with a high-quality morphological reconstruction of a real CA1 pyramidal cell, complete with detailed soma and dendrites, and added axonal compartments from a separate reconstruction. They then curated a new set of ion channel models, carefully re-evaluating each one against the latest experimental data on channel kinetics and distribution. The final model contains thirteen distinct active voltage- and calcium-gated ion channels, including sodium channels with different kinetics for the soma, axon, and dendrites; multiple types of potassium channels; and four types of calcium channels. Calcium currents were described using the Goldman-Hodgkin-Katz equation, which provides a better fit for calcium-mediated currents than the traditional conductance-based formalism.

Parameters that were not sufficiently constrained by experimental data were determined through automated optimization using the Neuroptimus tool, which employs the Covariance Matrix Adaptation Evolution Strategy algorithm. The researchers ran twenty independent optimizations simultaneously from different random starting points to avoid getting stuck in local minima. The optimization targeted electrophysiological features extracted from somatic whole-cell recordings of real rat CA1 pyramidal cells, including spike counts, interspike intervals, action potential shapes, and voltage responses to both positive and negative current injections. When a parameter consistently drifted toward the edge of its allowed range, the team expanded the boundary and re-ran the search.

Validation was equally rigorous. The team used all five built-in tests of the HippoUnit framework, which automatically simulates experimental protocols and compares model behavior against experimental data. The Somatic Features Test examined voltage responses to current injections; the Backpropagating Action Potential Test evaluated how action potentials propagate from the soma into the dendrites; the Post-synaptic Potential Attenuation Test measured how synaptic signals diminish as they travel toward the soma; the Depolarization Block Test probed the cell’s response to strong sustained stimulation; and the Oblique Integration Test assessed nonlinear synaptic integration in thin dendritic branches. The models performed well on four of the five tests, with the depolarization block test revealing an interesting discrepancy that the researchers traced to the simplified axonal morphology in their model.

One of the most consequential findings concerned the NMDA receptor, a type of glutamate receptor critical for nonlinear dendritic integration. The team systematically evaluated eight different published models of NMDA receptor voltage dependence and found that none produced fully satisfactory results in their cell model. Some showed far too much supralinear summation of synaptic inputs, others remained sublinear, and none matched the experimental observations of how clustered synaptic inputs on oblique dendrites produce dendritic spikes. By systematically tuning the parameters of a sigmoid function describing the voltage-dependent magnesium block of the receptor, they identified a voltage dependence that matched experimental data both qualitatively and quantitatively. The result suggests that the precise voltage dependence of NMDA receptors has a major impact on how CA1 pyramidal cells integrate their inputs.

A second major contribution concerned dendritic spines, the tiny protrusions that receive most excitatory synaptic input. Modeling all 29,602 spines explicitly increases computational cost by more than an order of magnitude, so the team compared three approaches: merging spine properties into the parent dendrite using a surface-correction factor, explicitly modeling only the spines that receive synaptic input, and explicitly modeling every spine. Remarkably, for somatic current responses, backpropagating action potentials, and synaptic potential attenuation, all three versions behaved almost identically. But for nonlinear synaptic integration in oblique dendrites, the version without explicit spines produced clearly different results, because the high electrical resistance of the spine neck amplifies local voltage transients and strongly activates voltage-dependent mechanisms in the spine head. The hybrid approach—explicitly modeling only spines that receive synapses—offered a practical compromise.

The researchers also systematically blocked each ion channel in the model and compared the effects against what is known from experimental pharmacology. Blocking sodium channels prevented firing entirely; blocking A-type potassium channels eliminated the distance-dependent attenuation of backpropagating action potentials; blocking the hyperpolarization-activated current abolished the voltage sag during negative current injections. These channel-by-channel validations confirmed that the model’s components behave like their biological counterparts. The team also performed gradual perturbations, shifting each parameter by small amounts to probe how sensitive the model’s behavior is to precise parameter values, finding that a few critical parameters—particularly those governing sodium channels—had dominant effects across all tests.

The new model, along with all optimization scripts and validation results, has been made publicly available, and the researchers hope it will serve as a reliable component of detailed hippocampal network models as well as a platform for studying synaptic plasticity and dendritic computation. Just as importantly, the systematic workflow they developed—combining curated biophysical mechanisms, automated parameter fitting, and multi-paradigm validation—should generalize to other cell types. As the authors note, the most critical requirement for building faithful data-driven models of any neuron is the availability of comprehensive experimental data, and their approach provides a template for how such data can be transformed into models that are accurate, reusable, and broadly applicable across the computational neuroscience community.

Subject of Research: Development and validation of a detailed biophysical model of hippocampal CA1 pyramidal neurons

Article Title: How to train your neuron: developing a detailed, up-to-date, multipurpose model of hippocampal CA1 pyramidal cells

Article References: Tar, L., Sáray, S., Mohácsi, M., Freund, T. F., & Káli, S. (2026). How to train your neuron: developing a detailed, up-to-date, multipurpose model of hippocampal CA1 pyramidal cells. Frontiers in Computational Neuroscience, 20, Article 1897214. https://doi.org/10.3389/fncom.2026.1897214

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1897214

Keywords: CA1 pyramidal cells, hippocampus, computational neuroscience, biophysical modeling, ion channels, dendritic spines, NMDA receptors, Neuroptimus, HippoUnit, parameter optimization, synaptic integration, dendritic computation

The model, developed by Chun-Hui Huang, Wen-Shuo Chen, and colleagues at National Taiwan University and National Taiwan Ocean University, is built on the classic Hodgkin-Huxley formalism, the biophysical framework that describes how voltage-gated ion channels generate action potentials. Rather than lumping brain regions into abstract nodes, the testbed resolves eleven distinct neuronal populations spanning cortex, thalamus, striatum, and the basal ganglia: cortical pyramidal cells, two classes of cortical interneurons, thalamocortical relay cells, reticular thalamic neurons, striatal fast-spiking interneurons, direct- and indirect-pathway medium spiny neurons, neurons of the external and internal globus pallidus, and subthalamic nucleus cells. Each population carries its own complement of sodium, potassium, calcium, and adaptation currents, allowing the simulation to reproduce complex discharge patterns such as the post-inhibitory rebound bursting that emerges from T-type calcium channels in the subthalamic nucleus and thalamus.

To emulate Parkinson’s disease, the researchers did not simply dial down a single parameter. Instead, they reconfigured the network to mimic the circuit-level consequences of dopamine depletion: cortical drive to the direct pathway was weakened, input to the indirect pathway was strengthened, and feedforward inhibition within the striatum was reduced. Background currents to the globus pallidus and subthalamic nucleus were also adjusted, with the subthalamic baseline dropping from 33 to 23 microamps. The calibrated parkinsonian state showed hallmark reorganization: beta-band power, oscillations in the 13 to 30 hertz range associated with rigidity and slowness of movement, rose nearly fourfold in the internal globus pallidus and fourfold in indirect-pathway striatal neurons, while firing rates of the two medium spiny neuron populations shifted dramatically upward.

Against this pathological backdrop, the team ran two interventions on identical networks. The first was conventional: 130-hertz deep brain stimulation of the subthalamic nucleus, delivered as symmetric, charge-balanced biphasic pulses with a pulse width of 0.06 milliseconds per phase. The second was an idealized periodic visual input, generated by a prescribed input unit that delivers rhythmic pulse trains directly to cortical pyramidal cells and thalamic relay cells. The authors are careful to note that this input unit is not a biophysical model of the retina or the lateral geniculate nucleus; it bypasses the sensory filtering that real visual signals undergo, and the archived identifier suggesting a 40-hertz condition is a legacy label—the implemented drive is a 10-hertz base-pulse train, with the 38 to 42 hertz band analyzed purely as an output measure.

The contrast between the two interventions was stark. Deep brain stimulation crushed the model’s internal measure of subthalamic beta-band synaptic current, reducing it from 0.0184 to 0.0018 arbitrary units—a suppression of roughly ninety percent, with a Cohen’s d effect size of 3.9 and a Holm-corrected p-value of 1.1 times ten to the minus fifth power. The periodic visual input did the opposite on this measure: subthalamic beta power actually rose by more than double. But the visual input produced something deep brain stimulation did not. Thalamic relay cell firing jumped from 8.6 to 35.3 hertz, previously silent reticular thalamic neurons woke up to 6.56 hertz, and cortical pyramidal firing nearly doubled from 4.86 to 10.33 hertz, accompanied by significant increases in broadband gamma power and narrowband power in the 38 to 42 hertz window. Deep brain stimulation, by contrast, left cortical firing statistically unchanged.

One of the most intriguing findings concerns what happened to the thalamus under deep brain stimulation. Spontaneous firing of thalamocortical relay cells fell to zero at every tested stimulation amplitude from 1 to 5 milliamps per square centimeter—a result that could, at first glance, suggest the stimulation had silenced the thalamus entirely, acting like a reversible lesion. The team dug deeper. In a full-resolution diagnostic run, they inspected 2.5 million membrane voltage samples from all ten modeled relay cells and found every sample finite and subthreshold, ranging from about -71 to -58 millivolts, with no crossings through zero millivolts. That pattern is inconsistent with sustained depolarization block, the pathological state in which neurons are stuck in a hyperpolarized inactivity. Instead, the data pointed to subthreshold quiescence: the relay cells remained alive and responsive, simply not firing spontaneously under the calibrated balance of excitatory and inhibitory inputs.

To test whether this quiescence depended on one particular parameter, the researchers ran a matched-seed sensitivity analysis on the inhibitory coupling from the internal globus pallidus to the thalamus, varying it by twenty percent in either direction. Across all thirty runs, thalamic spontaneous firing remained at zero, and all 750,000 saved voltage samples stayed subthreshold, between -72 and -55 millivolts. A further three-parameter sensitivity analysis—varying a GABA decay constant, the subthalamic-pallidal coupling conductance, and the subthalamic injected current, each by twenty percent—left all four principal outcomes intact, with beta suppression under deep brain stimulation remaining between 88.9 and 95.8 percent. These checks do not prove the model is robust to every conceivable perturbation, but they show the headline results are not artifacts of a single finely tuned value.

The dose-response story added another layer of nuance. Rather than a clean monotonic relationship between stimulation intensity and beta suppression, the amplitude series revealed a non-monotonic response: at 1 milliamp per square centimeter, subthalamic beta actually increased above the unstimulated parkinsonian baseline, while at 2 through 5 milliamps the mean beta values settled into a narrow band corresponding to over ninety percent suppression. Because no amplitudes between zero and one milliamp were sampled, the exact transition point where thalamic firing collapses remains unresolved. The authors also stress that the model’s current density is not directly translatable to clinical device current, which depends on electrode geometry, volume conduction, and axonal recruitment in real tissue.

The study is equally candid about its boundaries. The subthalamic beta measure is derived from outgoing synaptic current rather than a simulated local field potential, and the model omits the hyper-direct cortico-subthalamic pathway now considered a key driver of beta oscillations in human Parkinson’s disease—perhaps explaining why the canonical parkinsonian rise in subthalamic beta appeared in the pallidum and striatum instead. The simplified non-REM sleep representation failed to reproduce the pathological beta that clinical recordings show can persist during sleep, and the cortical module, with its simplified horizontal inhibition, produced stronger high-beta synchrony than would be expected clinically. The authors frame these as defined targets for future iterations rather than fatal flaws.

What emerges is not a treatment recommendation but a set of sharply defined, testable hypotheses. Deep brain stimulation and rhythmic visual input, in this calibrated network, produce fundamentally different response profiles: one suppresses pathological subthalamic beta while leaving cortical output largely untouched, the other recruits thalamic and cortical activity vigorously without touching the beta measure. Whether visual recruitment preserves information processing, whether thalamic quiescence under stimulation is compatible with faithful relay, and whether either profile translates to motor benefit in patients are questions the model cannot answer. But by placing both interventions on the same digital substrate—complete with wakefulness, sleep, and anesthesia states—the testbed gives experimentalists a concrete map of where to look next, and a reminder that in the brain, suppressing one circuit’s noise and amplifying another’s signal are not the same thing.

Subject of Research: Computational modeling of deep brain stimulation versus periodic visual input in the Parkinsonian cortico-basal ganglia-thalamic circuit

Article Title: An in silico cortico-basal ganglia-thalamic testbed for comparing visual and deep brain stimulation in Parkinson’s disease

Article References: Huang, C.-H., Chen, W.-S., Lai, J.-R., Cho, C.-H., Sung, Y.-T., Chang, C.-T., & Lin, C.-W. (2026). An in silico cortico-basal ganglia-thalamic testbed for comparing visual and deep brain stimulation in Parkinson’s disease. Frontiers in Computational Neuroscience, 20, Article 1950516. https://doi.org/10.3389/fncom.2026.1950516

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1950516

Keywords: Parkinson's disease, deep brain stimulation, computational neuroscience, Hodgkin-Huxley model, basal ganglia, thalamus, beta oscillations, visual stimulation, neuromodulation, cortico-basal ganglia-thalamic loop, in silico modeling, gamma oscillations

Myelin is built by oligodendrocytes, specialized glial cells that each form multiple sheaths around several axons. During development, oligodendrocyte precursor cells migrate through the central nervous system and mature, generating a dynamic period of sheath formation and loss followed by pruning. Neural activity shapes this process: increased firing enhances precursor proliferation and maturation, boosting myelination. For a long time, mature myelin was considered essentially stable, changing only when new oligodendrocytes appeared or in disease. More recent experimental evidence has overturned that view, showing that myelination remains dynamic throughout adulthood. Studies in the rodent cortex reveal that learning, experience, and behavior alter sheath length, number, thickness, and organization, and there is even growing evidence for activity-dependent myelin retraction. Neurons and oligodendrocytes form a bidirectional feedback loop that shapes the myelination program of the brain in real time.

The functional significance of this loop is profound. By remodeling myelin, oligodendrocytes tune axonal conduction velocity and thereby the delays with which action potentials arrive at their postsynaptic targets. This introduces a timing-based plasticity mechanism that operates in parallel to changes in synaptic strength, supporting learning, memory, and sensory and motor function. In the healthy brain, ADM optimizes and coordinates neural communication. But abnormal activity can hijack it. In certain forms of epilepsy, seizure activity promotes maladaptive hypermyelination that reinforces seizure patterns and drives disease progression, a feedback loop of dysfunction. In demyelinating disorders such as multiple sclerosis, compromised myelination underlies the sensory, motor, and cognitive deficits patients experience. Despite this significance, ADM lacks the strong theoretical framework that synaptic plasticity enjoys, largely because it involves multiple mechanisms interacting across widely different time scales: action potentials propagate in milliseconds, whereas myelin changes unfold over hours to days, and the relevant processes span subcellular signaling, single axons, gray matter microcircuits, and large-scale white matter networks.

Nguyen and Lefebvre organize the emerging computational literature into three classes of models, each operating at a distinct spatial scale. The first class focuses on individual axons, where neuron-glia interactions occur, and examines how changes in myelin structure map onto changes in conduction timing. Because structural remodeling in vivo often alters multiple myelin properties simultaneously, computational models can isolate the impact of individual variables and determine where causal relationships exist. These models are often biophysically detailed, explicitly representing action potential dynamics along myelinated axons composed of sheath and node compartments, frequently relying on Hodgkin-Huxley-like kinetics to simulate realistic ionic currents.

A landmark example comes from combined experimental-computational work by Ford and colleagues, who fit their model to geometrical measurements and electrophysiological recordings from gerbil auditory fibers. Building on the Halter and Clark multi-compartment model of the myelinated axon, which independently models intra-axonal, peri-axonal, and extra-axonal potentials, they established how the ratio between internode length and axon diameter influences conduction velocity. The smaller ratio observed in low-frequency auditory fibers alone accounts for a 15 percent increase in conduction velocity, and both axon types exhibit an optimal ratio that maximizes speed: internodes that are too short lack overall myelination, while depolarizations are not efficiently transferred between nodes when internodes are too long. Follow-up modeling of node length, fit to rat optic nerve and cortex measurements, revealed that with a constant number of ion channels per node, increasing node length almost linearly decreases conduction velocity due to increased capacitance and axial resistance, yet with constant channel density there is an optimal node length. Micrometer-scale changes in node length can increase conduction velocity by up to 20 percent in the optic nerve, and nodal remodeling presents a more metabolically efficient strategy for tuning conduction than altering internode architecture.

Axonal models have also examined how spatial heterogeneity in myelin coverage impacts signaling, injury susceptibility, and recovery. Myelin sheaths vary in length and are often non-uniformly arranged along axons, a patchy organization that is difficult to precisely manipulate in vivo but easy to control in simulation. Work by Scurfield and Latimer showed that, assuming the same number of remyelinated segments, transitions between short and long internodes slow conduction more than if all shorter internodes occupy one half of the axon and all longer ones the other. More recent modeling by Talidou and Lefebvre, using mouse imaging data to generate physiologically realistic sheath lengths and arrangements, demonstrated that axons with uniform myelin show negligible variability in conduction delays, whereas heterogeneous motifs show high, nonlinear, path-dependent variability, and that spatial heterogeneity increases the likelihood of conduction failure, making some motifs more vulnerable to demyelination. The major limitation of this class of models is that they do not explicitly model ADM; they treat myelin microstructure as a static parameter and model different myelination states as a proxy for its outcome. Still, they provide evidence that specific structural configurations can optimize spike timing, which is a primary goal of ADM.

The second class operates at the microcircuit level, where conduction velocity serves as a proxy for myelination and the feedback between neural activity and myelin adaptation is modeled directly. Talidou and colleagues built a recurrent network of excitatory spiking neurons in which conduction delays are state-dependent, using a Hebbian, use-dependent rule in which conduction velocity scales with firing rate along a given axon. Simulations showed that ADM acts as a homeostatic mechanism of gain control, temporally coordinating action potentials and up-regulating correlated activity, and that it can bidirectionally and reversibly alter conduction velocities in response to external stimuli, suggesting myelin plasticity responds dynamically to different brain states such as learning versus rest. Because synaptic connectivity was fixed, the results highlight how timing plasticity can complement synaptic plasticity in maintaining optimal brain function.

Pajevic and colleagues took a step further by asking how oligodendrocytes might integrate local spike timing as feedback to tune conduction velocities. In their framework, each oligodendrocyte releases a transient global intracellular signaling factor in response to action potentials propagating along any of the multiple axons it myelinates, while each axon carries a unique local myelination factor whose release is proportional to the global signal. This combination of global and local signals allows selective, bidirectional modulation of myelination along each axon. Model simulations reveal that oligodendrocytes can actively tune conduction velocities to selectively synchronize correlated spike trains while leaving uncorrelated trains unaffected, with synchronization greatest when firing rates are 10 Hz or less and transient responses span 10 to 40 milliseconds. This proposes a biologically plausible mechanism in which oligodendrocytes act as coincidence detectors of neural activity. The limitations are clear, however: existing microcircuit models rely on phenomenological learning rules, bypass explicit myelin structure, consider only excitatory networks, and compress the timescales of myelination for computational tractability.

The third class addresses the whole brain. Most inter-regional traffic flows through white matter, a heavily myelinated fiber network occupying roughly half of the mammalian brain volume, and large-scale models adopt phenomenological neural mass formulations that link directly to neuroimaging data such as fMRI, EEG, and MEG. Pajevic and colleagues emphasized that myelin-dependent changes in conduction velocity alter time delays, which can strongly influence brain oscillations and synchrony, drawing on the phenomenon of amplitude death, whereby dispersed delays suppress oscillatory activity. Using time-delayed Kuramoto oscillator networks, they showed that millisecond-scale changes in delays can impede the ability of neural assemblies to synchronize, and that delays directly alter oscillation frequency. Noori and colleagues then implemented a phenomenological learning rule in which conduction velocity along each axon changes according to the oscillatory phase difference between nodes, informed by macaque white matter structural connectivity. Their simulations revealed that ADM enables an initially incoherent network to self-regulate conduction velocities and achieve a globally phase-locked state, compensating for differences in tract lengths to maintain similar delays, and that plastic velocities enable synchronization at lower synaptic strengths and across wider frequency bands, even restoring synchrony after perturbations or injury. Park and Lefebvre confirmed mathematically that phase-dependent delays increase the stability of globally synchronous states and promote resilience to the loss of axonal connections.

Toward a unified framework, Nguyen and Lefebvre argue that the field remains fragmented and that future progress depends on interdisciplinary iteration between models and experiments. The physiological mechanisms underlying ADM remain unclear, and current models change myelination based on firing rate, individual spike timings, or synchrony, while ignoring oligodendrocyte heterogeneity and the contributions of astrocytes, precursor cells, and microglia. A particularly promising mechanistic candidate is calcium-dependent signaling: recent in vivo imaging shows that increases in intracellular calcium transients within developing oligodendrocytes trigger myelin sheath elongation, while sheath shortening is associated with lower transient frequency, suggesting a calcium-mediated feedback loop that could temporally bridge fast neural activity and slower myelin changes. Multiscale modeling offers a path across spatial scales, as illustrated by recent work linking a biophysically detailed axonal model of rhesus monkey prefrontal neurons to a spiking network showing that demyelination-induced propagation failures impair working memory but can be restored with remyelination. The authors conclude that future models should include dynamic conduction delays, analogous to the widespread adoption of variable synaptic weights, and that, like synaptic plasticity models, ADM models may clarify this form of plasticity in healthy and pathological brain states while inspiring new applications in engineering, machine learning, and artificial intelligence.

Subject of Research: Neuroscience

Article Title: A computational roadmap to activity-dependent myelination

Article References: Nguyen, T., & Lefebvre, J. (2026). A computational roadmap to activity-dependent myelination. Frontiers in Computational Neuroscience, 20, Article 1895171. https://doi.org/10.3389/fncom.2026.1895171

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1895171

Keywords: activity-dependent myelination, myelin plasticity, oligodendrocytes, conduction velocity, neural synchrony, computational neuroscience, white matter, brain plasticity, computational, roadmap, activity-dependent, myelination

Electroencephalography, or EEG, remains the workhorse of epilepsy diagnosis. The technique records the brain’s tiny electrical fields, typically just 10 to 100 microvolts in amplitude, through electrodes on the scalp, and offers the millisecond-level temporal resolution that brain imaging tools such as MRI cannot match. Yet the same signals are notoriously difficult for machines to interpret. They are non-stationary, meaning their statistical properties shift over time, they are non-linear, and they carry a low signal-to-noise ratio contaminated by eye movements, muscle activity and electrode drift. Manual review by a trained neurologist is slow, subjective and vulnerable to fatigue-induced error.

Earlier automated approaches relied on handcrafted features such as power spectral density, wavelet coefficients or common spatial patterns, fed into shallow classifiers like support vector machines and linear discriminant analysis. These pipelines demand task-specific engineering and struggle to capture the full complexity of EEG dynamics. Deep learning changed the calculus: convolutional neural networks, or CNNs, learn local waveform patterns directly from raw signals, while recurrent architectures such as Long Short-Term Memory networks track how those patterns evolve across time. But each approach has blind spots. CNNs alone have no notion of temporal context, and recurrent networks process sequences inefficiently and look in only one direction.

The new framework, dubbed CNN-BiLSTM-MHSA, addresses those gaps by stacking three modules into a single end-to-end trainable network. A multilayer one-dimensional CNN front end, with filters of 32, 64 and 128, extracts localized morphological signatures such as spikes and sharp waves. A Bidirectional LSTM layer with 128 units then reads the feature sequence in both forward and backward directions, capturing how ictal activity unfolds and recedes over time. Finally, a Multi-Head Self-Attention mechanism with four heads adaptively weights the most discriminative moments in the recording, no matter how far apart they lie in time, and refines the recurrent representation through a residual connection and layer normalization.

The attention module serves a second purpose beyond raw accuracy: interpretability. Because attention weights can be visualized as temporal salience maps, clinicians can in principle see which segments of an EEG recording drove the model’s decision. The authors report that high attention values align with informative temporal regions, while low-weight regions correspond to comparatively irrelevant signal segments, a property that becomes increasingly important as such tools move toward clinical translation.

Before any of this learning took place, the raw data passed through a disciplined preprocessing pipeline. A fifth-order Butterworth bandpass filter with cutoffs at 0.5 and 40 Hz isolated the clinically relevant delta-through-gamma frequency range where epileptiform activity is typically expressed, and zero-phase forward-backward filtering eliminated the timing distortions that ordinary filters introduce. Each one-second epoch was then normalized to zero mean and unit variance so the network would learn morphological patterns rather than amplitude artifacts. To stretch a moderate dataset further, the team generated synthetic training variants by injecting Gaussian noise with a standard deviation of 0.01 and applying random amplitude scaling, emulating natural variability across patients.

The evaluation used the UCI/Kaggle Epileptic Seizure Recognition dataset, comprising 11,500 EEG segments of 178 samples each, sampled at 178 Hz and balanced across five classes: seizure activity, tumor-region activity, healthy brain region from tumor patients, and eyes-closed and eyes-open resting states in healthy subjects. With a stratified 80:20 train-test split, the framework achieved 98.83 percent accuracy, 98.95 percent precision, 98.84 percent recall, a 98.90 percent F1-score, 99.71 percent specificity, 99.80 percent ROC-AUC and 99.72 percent PR-AUC. One class achieved perfect recognition, and per-class accuracies exceeded 99 percent for most categories. On the efficiency side, the model contains just 494,483 parameters and classified samples at roughly 2.26 milliseconds each, a throughput of about 442 samples per second.

An ablation study confirmed that each component earns its place. A CNN alone reached 94.72 percent accuracy, adding the BiLSTM lifted that to 96.84 percent, and pairing the CNN with attention alone produced 97.41 percent. The full three-module system with data augmentation reached 98.83 percent, while the same architecture without augmentation managed 97.92 percent, showing that the augmentation strategy alone contributed nearly a full percentage point. The complete system also edged out a prior BiLSTM-based approach on the same dataset that had reported 98.70 percent accuracy.

Visualization analyses reinforced the picture of a model that has genuinely learned the structure of brain activity. Projections of the learned features into two dimensions using t-SNE revealed five well-separated clusters with minimal overlap, and intermediate convolutional feature maps showed strong responses around informative EEG regions while suppressing redundant signal components. The handful of misclassified recordings, the authors note, exhibited waveform characteristics inherently similar to neighboring categories, underscoring the difficulty of the residual cases rather than any systematic weakness in the architecture.

The researchers caution that clinical deployment will require external validation on diverse patient populations, along with attention to interpretability, data privacy, computational constraints and integration with existing hospital workflows. Their planned next steps include testing on larger clinical EEG corpora such as CHB-MIT and TUH EEG, developing lightweight versions for edge devices, and incorporating explainable AI techniques to support clinical decision-making. If those hurdles are cleared, an attention-guided network that reads a second of brain activity in about two milliseconds could bring automated, near-real-time seizure detection closer to the bedside.

Subject of Research: A hybrid deep learning framework using CNN, BiLSTM and multi-head self-attention for multiclass epileptic seizure recognition from EEG signals

Article Title: An attention-enhanced hybrid CNN-BiLSTM framework for multiclass epileptic seizure recognition using EEG signals

Article References: Kaur, B., Middha, S., Ali, A., Khan, N., Alhejaili, R., & Almutairi, A. H. (2026). An attention-enhanced hybrid CNN-BiLSTM framework for multiclass epileptic seizure recognition using EEG signals. Frontiers in Computational Neuroscience, 20, Article 1949040. https://doi.org/10.3389/fncom.2026.1949040

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1949040

Keywords: EEG, epilepsy, seizure detection, deep learning, CNN, BiLSTM, self-attention, neural networks, brain-computer interface, signal processing, computational neuroscience, machine learning

Ion channels are the protein macromolecules that conduct electrically charged ions across neuronal membranes, and their conductance is controlled by their conformational state. Markov models of molecular kinetics sit at the very foundation of neuroscience, describing the physical components that implement the brain’s computations. They were first applied to neuroscience to describe the ion channels responsible for generating action potentials, and because brains are composed of trillions of repeated components, computational neuroscience has what the study’s author calls a voracious appetite for these models. Yet despite their biological fidelity, Markov kinetic models have remained underutilized because they are notoriously difficult to simulate efficiently.

The mathematical core of the problem is the master equation, which describes how a probability distribution over the channel’s conformational states evolves over time. This initial value problem has an analytic solution expressed through the matrix exponential function, yielding what is known as the propagator matrix. The propagator matrix advances the state of the model over each time step, but computing it exactly is prohibitively expensive for large-scale simulations. Direct evaluation of the Taylor series expansion converges slowly and is numerically unstable, while better methods such as scaling and squaring with Padé approximants still involve matrix-matrix multiplications whose time complexity scales cubically with the number of model states.

Established numerical solvers fall into two camps, each with serious drawbacks. Explicit integration methods, such as the forward Euler and Runge–Kutta–Chebyshev approaches, evaluate the derivative and run the state forward in time. They are simple and efficient, with quadratic time complexity, but they can be numerically unstable, allowing small errors to accumulate and compound exponentially into sudden catastrophic failure. Implicit methods, including the backward Euler and Crank–Nicolson schemes, evaluate the derivative at the end of the time step and offer accurate, stable integration, but they require solving a system of equations via Gaussian elimination, giving them cubic time complexity. The NEURON simulator, the state-of-the-art tool for conductance-based neuronal modeling, uses implicit methods by default.

The AME method takes a fundamentally different approach. Instead of computing the propagator matrix from scratch at every time step, it constructs a piecewise polynomial approximation of the propagator matrix function. The input space is partitioned into a regular grid of non-overlapping regions, each containing a polynomial approximation fitted to sampled data using linear least squares regression. Within each region, the input vector is transformed into a normalized range, and because the function being approximated returns a full matrix, the method builds a separate approximation for every cell of the propagator matrix, with all approximations sharing the same partitions and polynomial form for computational efficiency.

What makes the method particularly elegant is how it measures and controls its own accuracy. The integration error is defined as the maximum absolute error in any component of the state vector after integrating over one unit of time, across all possible initial states and inputs. Exploiting the Chapman–Kolmogorov equation, the method squares propagator matrices to extend sample integration periods from one time step to a full unit of time. Because Markov models are linear and probability is a conserved quantity, the maximum error occurs when the entire probability mass sits in a single state, which means the error of the standard basis vectors bounds the error of all possible state vectors. A greedy algorithm then adjusts the number of input partitions until the residual integration error falls below a user-specified error target, and a hill-climbing algorithm searches the space of polynomial forms to maximize throughput.

The method also handles a subtlety specific to neurobiology: chemical concentrations such as neurotransmitter levels span many orders of magnitude, which is difficult to approximate with fixed-size piecewise regions. The AME method preconditions concentration inputs with a base-2 logarithmic transform and a small positive offset, determined experimentally by testing 50 offsets in a geometric progression and selecting the one that minimizes the root-mean-square residual error. The base-2 logarithm is preferred because efficient algorithms and even dedicated hardware instructions, such as the lg2 instruction in Nvidia PTX assembly, can compute it quickly.

To validate the approach, the author assembled a benchmark of 99 Markov models drawn from four independent sources, covering AMPA and NMDA receptors, all nine subtypes of voltage-gated sodium channels, and 88 models of voltage-gated potassium channels built from seven kinetic schemes. The models were obtained from ModelDB and GitHub, and some numerically unstable models were detected and rejected by the tooling. Accuracy comparisons using the analytic solution as ground truth showed that the backward Euler method is only first-order accurate, with error directly proportional to the time step, whereas the AME method controls its accuracy independently of the time step, achieving an error target of 10 to the power of minus 3 for all state and input vectors tested.

The practical consequences were demonstrated in simulations of action potentials, the electrical signals that broadcast information throughout the brain, where timing errors quickly cascade. A single-compartment neuron containing eight ion channels from the benchmark was stimulated to produce trains of action potentials. With the backward Euler method, action potentials were delayed in proportion to the time step, while simulations using the exact matrix exponential and the AME method delayed action potentials only in proportion to the square of the time step, limited by NEURON’s second-order cable solver. Varying the AME error target showed convergence by 10 to the power of minus 3, while the loosest target of 10 to the power of minus 1 caused simulations to fail to produce action potentials altogether. In speed benchmarks with 1,000 single-compartment neurons, the AME method consistently ranked fastest, followed by backward Euler and then the exact matrix exponential.

The accompanying open-source program, called ame and available under the MIT license on the Python Package Index, integrates the method into the NEURON simulator, and for comparison the exact analytic solution was also added to NEURON as a new solver option. On an NVIDIA B200 GPU, the method achieved integration speeds as low as 0.059 nanoseconds per step per instance for a four-state potassium channel model, corresponding to real-time simulation of roughly 420,000 instances. The method does have limitations: it is restricted to models with one or two input variables, its time complexity scales quadratically with the number of model states, and building the approximation can take up to an hour per model on state-of-the-art hardware. Nevertheless, because most voltage-gated ion channels genuinely have only one input and most phenomena of interest can be modeled with a dozen states or less, the method’s trade-offs are well suited to its purpose. By decoupling accuracy from the time step and delivering GPU-friendly, branch-free execution, the AME method paves the way for large-scale neuronal simulations to adopt far more realistic models of the ion channels that generate the brain’s electrical activity.

Subject of Research: A fast and accurate numerical method for simulating Markov kinetic models of ion channels in large-scale neuronal simulations

Article Title: Approximate matrix exponential: a fast and accurate solver for Markov kinetic models of ion channels

Article References: McDougall, D. A. (2026). Approximate matrix exponential: a fast and accurate solver for Markov kinetic models of ion channels. Frontiers in Computational Neuroscience, 20, Article 1901593. https://doi.org/10.3389/fncom.2026.1901593

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1901593

Keywords: ion channels, Markov kinetic models, matrix exponential, NEURON simulator, GPU computing, numerical integration, computational neuroscience, propagator matrix, action potentials, conductance-based models, piecewise polynomial approximation, neuronal simulation