Front Comput Neurosci [JOURNAL]
Sun
200 papers
Learning under constraints: a theoretical framework for comparing resource-constrained learning in biological and artificial systems.
Kucherov D, Dolgikh S, Podelskyi S
Front Comput Neurosci · 2026 · PMID 42388384 · Full text
Learning processes, cognitive architectures, available resources, and methods for sampling the environment and generating intelligent responses in complex sensory domains can differ significantly between natural and arti...
Learning processes, cognitive architectures, available resources, and methods for sampling the environment and generating intelligent responses in complex sensory domains can differ significantly between natural and artificial systems. In this work, we present theoretical and modeling-based analysis of early-stage learning under resource constraints, comparing biological intelligence with a class of freely evolving, weakly constrained artificial systems (FEW), focusing on essential resource constraints such as computational capacity, memory, and energy. We develop quantitative models of sensory exploration and learning under strong and weak resource constraints, formalizing how limitations in energy, memory, and computational capacity shape sampling strategies and learning dynamics. For biologically constrained systems, we show that steep anisotropy in the cognitive cost gradient induces prioritized, depth-oriented exploration within limited sensory regions, leading to robust and resource-efficient learning. In contrast, we demonstrate that FEW systems, despite access to abundant resources, face a paradox of unconstrained learning: in the absence of intrinsic prioritization and evaluative feedback, uniform or random sampling leads to inefficient exploration of the sensory domain. To examine this challenge, we introduce a comparative framework for evaluating sensory traversal strategies and show that no single strategy dominates across prioritization accuracy, robustness, and resource efficiency. Instead, our analysis suggests a meta-strategy approach, in which adaptive selection among exploration strategies optimizes empirical success while preserving empirical accountability required for adaptive optimization. These results clarify the functional role of constraints in biological learning and provide principled guidance for the design of next-generation artificial learning systems operating in complex sensory environments.
MsGCN: a multi-stream graph convolutional network for multiband PLV graph fusion in EEG-based biometric identification.
Tian W, Yang J, Ju X … +2 more , Li M, Hu D
Front Comput Neurosci · 2026 · PMID 42388383 · Full text
INTRODUCTION: EEG-based biometric identification has attracted extensive attention due to its high security and uniqueness. Functional connectivity features derived from EEG exhibit strong individual specificity, yet exi...
INTRODUCTION: EEG-based biometric identification has attracted extensive attention due to its high security and uniqueness. Functional connectivity features derived from EEG exhibit strong individual specificity, yet existing methods do not fully leverage the complementary identity information contained in multiband functional connectivity features. METHODS: This study proposes a multi-stream graph convolutional network (MsGCN) for EEG-based biometric identification by fusing graph representations derived from multiband phase-locking value (PLV) matrices. The model processes PLV matrices from multiple frequency bands through parallel GCN branches and performs end-to-end identification using fully connected layers. Experiments on the public PhysioNet Motor Movement/Imagery dataset evaluated the method under non-preprocessed conditions, cross-task settings, channel reduction, and different graph binarization thresholds. RESULTS: MsGCN achieved 99.50% accuracy on preprocessed data and 98.12% on non-preprocessed data, showing numerically higher accuracy than the selected CNN and GCN baselines under the unified protocol. The model also showed improved robustness in cross-task identification, reduced-channel settings, and across a wide range of thresholds. DISCUSSION: These results suggest that multiband PLV graph fusion can improve robustness to preprocessing conditions, task variation, channel reduction, and threshold selection under the evaluated dataset and experimental settings.
AI-driven neuroanalytic modeling for mental health: multichannel CNN-based autism spectrum disorder detection via facial pattern analysis.
Kaur N, Singh P, Singh K … +5 more , Khan J, Hussain D, Gu YH, Aljuaidi R, Waheb Rajkhan N
Front Comput Neurosci · 2026 · PMID 42375820 · Full text
INTRODUCTION: Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by abnormal brain connections, impaired cognitive functions, and dysfunctional behaviors, which, in mental health, is a major c...
INTRODUCTION: Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by abnormal brain connections, impaired cognitive functions, and dysfunctional behaviors, which, in mental health, is a major challenge to diagnose at an early age. Recent developments in Artificial Intelligence (AI) and computational neuroscience have made it possible to use neuroanalytic methods to identify subtle patterns related to brain disorders. Inspired by this, this study investigates facial pattern analysis as a non-invasive surrogate biomarker. METHODS: A neuroanalytic deep-learning model is suggested on the basis of a Modified Histogram of Oriented Gradients-based Multichannel Convolutional Neural Network (MHMCNN). The technique comprises three steps, that is, (i) preprocessing and normalization of facial images, (ii) extraction of discriminative neuro-inspired features based on modified HOG descriptors, and (iii) multichannel CNN-based classification to discover complex structural and micro-pattern variations. The model is trained and tested on a publicly accessible facial autism dataset, and the performance of the model is tested using -fold cross-validation. RESULTS: The proposed MHMCNN framework achieved a validation accuracy of 98% and a test accuracy of 96.2%, demonstrating strong generalization capability for ASD facial image classification. The model attained a training accuracy of 99.8%, indicating effective feature learning during optimization. The combination of handcrafted feature descriptors and deep learning improves the feature representation and the strength of classification. Experimental findings support the enhanced generalization and stable recognition of ASD-related patterns. DISCUSSION: The results emphasize the possible application of AI and computational neuroscience in neuroanalytic pattern detection in mental health diagnostics. The proposed solution offers a cost-effective and scalable solution to early screening of ASD by allowing observable facial characteristics to be related to underlying neurodevelopmental features. The work has helped in filling the gap between the phenotypic observations and the diagnosis of the disorder of the brain. Future studies will target the use of multimodal integration of neuroimaging and behavioral data to enhance understanding and clinical utility. CONCLUSION: This research introduces a new combination of AI and neuroanalytic principles to detect ASD that can further advance computational neuroscience-based mental health diagnostics. The suggested framework offers a scalable and affordable outcome of early screening and future expansion to multimodal frameworks of neuroimaging and behavioral data to increase clinical utility and interpretation.
Modeling multiscale neural dynamics for EEG-based emotion recognition using an attentive wavelet-transformer framework.
Soundariya RS, Thangaraj P
Front Comput Neurosci · 2026 · PMID 42344184 · Full text
Electroencephalography (EEG)-based emotion recognition faces challenges such as signal noise, non-stationarity, inter-subject variability, and class imbalance, limiting its practical application in affective computing an...
Electroencephalography (EEG)-based emotion recognition faces challenges such as signal noise, non-stationarity, inter-subject variability, and class imbalance, limiting its practical application in affective computing and clinical diagnostics. This study introduces the Attentive Wavelet-Transformer Network (AWT-Net), a novel framework integrating Hierarchical Wavelet Packet Decomposition (HWPD), Empirical Wavelet Transform with Kalman filtering (EWT-Kalman), Multi-Head Self-Attention (MHSA), and a Hybrid Spatio-Temporal Transformer (HSTT) to address these issues. The proposed work is evaluated on a custom EEG dataset (2,132 samples, 14 channels, 28 subjects) and the DEAP dataset (1,280 trials, 40 channels, 32 subjects), AWT-Net achieves window-level, subject-dependent accuracy of 99.61% on DEAP and 99.34% on custom EEG. Under stricter evaluation protocols, accuracy is 99.30% (trial-wise grouped) and 97.23% (subject-independent LOSO) on DEAP, demonstrating robust generalization across varying validation conditions. Comparisons with baseline models (LSTM: 89.42%, CNN-LSTM: 91.75%) are provided under equivalent subject-dependent protocols, while LOSO comparisons (Elrefaiy et al.: >97.00%, Bagherzadeh et al.: ~77.75%) highlight cross-subject performance. Error rates are significantly reduced to 0.70% (EEG) and 0.39% (DEAP), compared to 6.69-10.58% for baselines, with statistical validation confirming large effect sizes (Cohen's d: 1.82-2.14). AWT-Net's adaptive focal loss mitigates class imbalance, while HWPD and EWT-Kalman enhance noise robustness. These results demonstrate AWT-Net's potential for real-time emotion recognition, advancing applications in healthcare and human-computer interaction.
New directions for complex systems in contemporary neuroscience: a morphodynamic and emergent function approach.
Oruro EM, Pardo GE, Rasia-Filho AA … +1 more , Garcia-Cairasco N
Front Comput Neurosci · 2026 · PMID 42328689 · Full text
We propose that current Neuroscience approaches can benefit from further integrating morphodynamics across different scales of brain organization and neural network emergent functions in complex systems. While emergence...
We propose that current Neuroscience approaches can benefit from further integrating morphodynamics across different scales of brain organization and neural network emergent functions in complex systems. While emergence in neuroscience is commonly addressed at higher organizational levels, here we consider neuronal morphology itself as an emergent level of organization. Progressing from form-based complexity views, early models of neuronal morphogenesis, and functional approaches, we integrate cell morphology to behavior with particular relevance to the following issues: (1) Neuronal Morphological Diversity and Circuitry Function, (2) Mother-Infant Relationships, and (3) Epilepsy and Neuropsychiatric Comorbidities. The structure of neurons and their connectivity within the brain volume are morphodynamic features that emerge from dynamic interactions among morphogenetic elements, the local cell neighborhood, and synaptic connections. In turn, the emergent functions of networks are organized around a series of conceptual, experimental, and computational foundations. Complex systems neuroscience combines such data with additional high- and multiscale information to develop models organized around structure, function, and behavioral displays in both normal and pathological conditions. Here, we present and discuss examples that approximate this framework, drawing on animal models and human data. Such an integrated approach aligns with the ongoing efforts promoted by UNESCO's "UniTwin Complex Systems Digital Campus" (CS-DC) to collaboratively address open, multiscale problems in neuroscience and complex systems.
NMDA receptor kinetics drive distinct routes to chaotic firing in pyramidal neurons.
Borjkhani M, Borjkhani H, Sharif MA … +2 more , Bahrami F, Janahmadi M
Front Comput Neurosci · 2026 · PMID 42318008 · Full text
INTRODUCTION: Neuronal firing patterns emerge from complex interactions between intrinsic membrane properties and synaptic receptor dynamics. N-methyl-D-aspartate (NMDA) receptors critically shape calcium influx and syna...
INTRODUCTION: Neuronal firing patterns emerge from complex interactions between intrinsic membrane properties and synaptic receptor dynamics. N-methyl-D-aspartate (NMDA) receptors critically shape calcium influx and synaptic plasticity through their voltage-dependent Mg block and prolonged activation kinetics, yet how their closing kinetics interact with glutamatergic drive and GABAergic modulation to control neuronal dynamics and information processing remains incompletely understood. METHODS: We developed a Hodgkin-Huxley-type computational model incorporating NMDA, AMPA, and GABA receptor kinetics to investigate how the NMDA receptor closing rate β and glutamatergic stimulation frequency control neuronal dynamics. We performed a systematic analysis of over 2.9 million inter-spike intervals across a large multi-parameter sweep of NMDA kinetics, glutamatergic stimulation frequency, and GABAergic modulation. Dynamical behavior was characterized using entropy-Lyapunov correlation analysis and frequency-dependent bifurcation analysis, and CaMKII phosphorylation was quantified to link kinetic regimes to downstream plasticity signaling. RESULTS: The analysis revealed two mechanistically distinct pathways to firing irregularity. Pathway 1 (rapid-deactivation irregularity) emerged under relatively fast NMDA deactivation combined with specific input-frequency conditions, producing deterministic chaos with compromised information encoding. Pathway 2 (prolonged-activation irregularity) resulted from slow NMDA deactivation under weak drive, creating irregularity through sustained receptor activation and calcium influx. An optimal kinetic window emerged at β = 0.042 ms, maximizing information transfer (0.275 bits) while maintaining stable dynamics. Entropy-Lyapunov correlation analysis confirmed deterministic chaos, and frequency-dependent bifurcation analysis demonstrated progressive narrowing and displacement of chaotic windows across the analyzed β range as stimulation frequency increased. GABAergic inhibition provided frequency-selective stabilization, expanding the stable parameter space by 34.2% while preserving gamma oscillations. CaMKII phosphorylation analysis revealed that prolonged NMDA activation maintained elevated phosphorylation levels, creating conditions for pathological long-term potentiation. DISCUSSION: These findings establish NMDA receptor kinetics as fundamental controllers of cortical excitability and information processing. The dual-pathway framework provides mechanistic insights into addiction-related memory formation, where prolonged NMDA activation enables pathological plasticity, and into visual processing disorders, where altered kinetics disrupt retinal function and cortical oscillatory balance. The identification of optimal kinetic windows and frequency-selective GABA modulation suggests therapeutic strategies based on kinetically specific interventions for neuropsychiatric disorders involving NMDA dysfunction.
Schumann-anchored golden ratio organization of human neural oscillations.
Lacy M
Front Comput Neurosci · 2026 · PMID 42312247 · Full text
INTRODUCTION: Human neural oscillations are organized according to golden ratio (φ = 1.618) mathematics: frequencies follow where ≈7.6 Hz. This architecture manifests as spectral peak depletion at integer positions (b...
INTRODUCTION: Human neural oscillations are organized according to golden ratio (φ = 1.618) mathematics: frequencies follow where ≈7.6 Hz. This architecture manifests as spectral peak depletion at integer positions (band boundaries) and enrichment at half-integer positions (band centers), providing empirical validation of a previously theorized φ architecture and identifying the absolute fundamental frequency = 7.6 Hz. This organization was discovered and validated through two complementary studies. STUDY 1—TRANSIENT EVENTS: Analysis of 1,366 Schumann Ignition Events (SIEs)-transient episodes of multi-band network synchronization at Earth-resonant frequencies-across 91 participants, 661 sessions, and three EEG devices characterized harmonic frequencies that suggested φ relationships (< 1% mean ratio error). Individual frequencies varied independently across events (all || < 0.03), yet ratio precision was preserved-an "independence-convergence paradox" indicating population-level rather than event-level constraints. Null controls confirmed genuine organization (Cohen's = 1.44, < 0.0001). STUDY 2—SINGLE-CHANNEL SPECTRAL ARCHITECTURE: Spectral parameterization of 244,955 oscillatory peaks across 968 sessions confirmed predictions derived from the φ framework: boundaries showed -18% depletion, attractors +21% enrichment, and noble positions (+0.618) +39% enrichment in aggregate cross-band analysis. The framework extends to an eight-position hierarchy including "inverse nobles" (+0.764, +0.854)-symmetric to regular nobles about the attractor-which inherit stability through multi-scale Fibonacci pathways. Gamma exhibited strongest aggregate adherence (+144.8% at Noble1 in cross-band analysis), consistent with functional requirements for precise phase relationships, though this aggregate figure may partially reflect cross-band density effects (see Section 6.8, Limitation 5). Independent replication in the EEGEmotions-27 dataset (612,990 peaks, 2,342 sessions) confirmed the same qualitative pattern with Kendall's τ = 1.0. SYNTHESIS: Two independent methodological approaches-transient event detection and single-channel spectral parameterization-converge on identical conclusions: neural oscillations follow φ organization with perfect position ordering (Kendall's τ = 1.0) across all analyses. The fundamental frequency = 7.6 Hz emerges independently from geophysical monitoring of Earth's Schumann Resonance and from neural spectral optimization, agreeing within 0.4%. These findings support a "substrate-ignition" model: the φ lattice exists continuously as an architectural scaffold organizing neural oscillations, while transient high-coherence events (SIEs) represent moments when this substrate is amplified and frequencies "snap" into tighter compliance. The golden ratio's unique mathematical properties-maximal resistance to mode-locking between frequency bands combined with precise Fibonacci-mediated cross-frequency coupling pathways-may represent evolution's solution to a fundamental computational challenge: maintaining independent parallel processing streams (segregation) while enabling flexible, controlled communication between them (integration).
Toward model-guided electrophysiology-Encoding of chirps in the electrosensory periphery of .
Barayeu A, Benda J, Grewe J
Front Comput Neurosci · 2026 · PMID 42293466 · Full text
Models formalize our understanding of a system and generate hypotheses that can be tested experimentally. In this study, we use a previously developed model of p-type electroreceptor afferents to support electrophysiolog...
Models formalize our understanding of a system and generate hypotheses that can be tested experimentally. In this study, we use a previously developed model of p-type electroreceptor afferents to support electrophysiological observations regarding the encoding of chirps in the electrosensory periphery of the weakly electric fish . These animals employ their self-generated quasi-sinusoidal electric fields to navigate, find prey, and communicate. Electrocommunication happens in an electrosensory context that is defined by the superposition of the electric fields of the interacting animals. Within this context, chirps-brief excursions of the electric organ discharge frequency-interrupt the periodic interference pattern and are encoded in the responses of the electroreceptor afferents. Behavioral observations highlighted the immediate importance of chirps happening in contexts that were believed to be far outside the electroreceptor tuning, i.e., not encoded at all. Combining experiments with modeling, we show that chirps are nevertheless encoded under these conditions, identify how the encoding works in such contexts, and provide a deeper understanding of chirp encoding in the electrosensory periphery.
Editorial: Cerebellar computations across the lifespan.
Oostland M, Verpeut JL
Front Comput Neurosci · 2026 · PMID 42293465 · Full text
Abstract loading — click title to view on PubMed.
Binarized neural networks converge toward algorithmic simplicity: empirical support for the learning-as-compression hypothesis.
Sakabe EY, Abrahão FS, Simões A … +4 more , Colombini E, Costa P, Gudwin R, Zenil H
Front Comput Neurosci · 2026 · PMID 42293464 · Full text
Understanding and controlling the complexity of neural networks is a central challenge in machine learning, with implications for generalization, optimization, and model capacity. While most approaches rely on entropy-ba...
Understanding and controlling the complexity of neural networks is a central challenge in machine learning, with implications for generalization, optimization, and model capacity. While most approaches rely on entropy-based loss functions and statistical metrics, these measures often fail to capture deeper, causally relevant algorithmic regularities embedded in network structure. We propose a shift toward algorithmic information theory, using binarized neural networks (BNNs) as a first proxy. Grounded in algorithmic probability (AP) and the universal distribution it defines, our approach characterizes learning dynamics through a formal, causally grounded lens. We apply the Block Decomposition Method (BDM), a scalable approximation of algorithmic complexity based on AP, and demonstrate that it more closely tracks structural changes during training than entropy, generally exhibiting stronger correlations with training loss across a wide range of architectures, datasets, and randomized training runs. These results support the view of training in BNNs as a process of algorithmic compression, where learning corresponds to the progressive internalization of structured regularities. In doing so, our work offers a principled estimate of learning progression and suggests a framework for complexity-aware learning and regularization, grounded in first principles from information theory, complexity, and computability.
Editorial: AI and neuroscience: integrating knowledge, reasoning, and theory of mind.
Edwards DJ, Zou B, Lowe R … +1 more , Owens A
Front Comput Neurosci · 2026 · PMID 42238304 · Full text
Abstract loading — click title to view on PubMed.
NEURONpyxl: fast, flexible, Python-integrated simulation of biophysical neural networks with complex plastic synapses.
Dickman U, Thomas PJ, Chiel HJ … +2 more , Byrne JH, Neveu CL
Front Comput Neurosci · 2026 · PMID 42238303 · Full text
INTRODUCTION: NEURON has been widely used as an empirically-based simulation tool, especially for multi-compartment conductance-based neuronal modeling. The network mediating feeding in Aplysia californica has been exten...
INTRODUCTION: NEURON has been widely used as an empirically-based simulation tool, especially for multi-compartment conductance-based neuronal modeling. The network mediating feeding in Aplysia californica has been extensively studied as a model central pattern generator. Understanding the relationship between network parameter values and their effect on animal behavior is of key importance in systems such as the Aplysia feeding apparatus, where detailed biophysical models can be constructed. OBJECTIVE: This study aims to develop a new Python tool called NEURONpyxl that reads parameters from a spreadsheet to construct full neural networks to make it easier to create complex models in the NEURON simulation environment, incorporating short-term forms of plasticity such as depression or facilitation. METHODS: Test simulations from well-understood networks were created in NEURONpyxl, and compared to simulation results of the same network in another neural simulator, the Simulator for Neural Networks and Action Potentials (SNNAP), which has previously been used to model conductance-based networks that include complex synaptic connections and multiple forms of synaptic plasticity. NEURONpyxl was then used to conduct a parameter grid search to optimize conductances in a previously developed network model of Aplysia feeding behavior. RESULTS: Simulations of the test networks in NEURONpyxl and SNNAP produced numerically equivalent results, with differences remaining within the expected margin of error arising from numerical integration and implementation details. We then located parameter values that generated simulated motor patterns with durations of protraction and retraction that matched biological feeding behavior under different mechanical loads. CONCLUSION: NEURONpyxl simplifies building and simulating complex neural networks with different forms of synaptic plasticity, and locating physiologically relevant parameter values. With NEURONpyxl, future work may include the creation of ensembles of network models and the integration of biomechanics with complex conductance-based networks.
Rehab-DRLX: explainable neurorehabilitation prognosis using deep reinforcement learning and transformer-based models.
Alsolai H, Khan S, Mahendran RK … +3 more , Panwar A, Alabduallah BI, Alhayan F
Front Comput Neurosci · 2026 · PMID 42232896 · Full text
Neurorehabilitation poses a crucial problem in clinical recovery tasks, particularly for individuals with poor motor functions and neurological impairments, and problems in activities of daily living (ADL). To resolve th...
Neurorehabilitation poses a crucial problem in clinical recovery tasks, particularly for individuals with poor motor functions and neurological impairments, and problems in activities of daily living (ADL). To resolve this, we design a novel model, Rehab-DRLX, with a hybrid deep learning (HDL) framework that combines deep reinforcement learning (DRL) with an explainable transformer model to provide interpretable, accurate prognostic results. The propounded model is designed to effectively process the multimodal data inputs, which include clinical records, sensor-entrenched motion data, and neuroimaging, along with time-dependent recovery patterns from its reinforced representation learning (RRL) module. The RRL module employs a convolutional neural network (CNN) within the DRL agent, which performs spatiotemporal feature encoding and dynamically recovers a policy from its reward-guided learning method. To ensure interpretability, the explainable prognosis transformer (XPT) is utilized, which contains clinical contextual positional encoding and a hierarchical attention mechanism to enable transparent and reliable decision-making. This duality in the Rehab-DRLX architecture enables effective forecasting of the recovery outcomes, including functional independence probability, with both interpretability and accuracy, addressing the drawbacks of conventional black box prognosis tools. The experimental results of Rehab-DRLX show the noteworthy improvements in metrics such as accuracy (94.6%), F1-score (0.93), root mean square (RMSE) (0.082), and mean absolute error (MAE) (0.061) compared to existing studies. The ablation studies reveal the significant contribution of every architectural component and its overall performance. The results show the practical viability of Rehab-DRLX, which not only improves decision-making but also builds clinical trust through explainable insights.
Deep learning guided propofol ketamine dosing and inflammation trajectories in elderly burns.
Yuan X, Wang G, Jiang X … +1 more , Miao W
Front Comput Neurosci · 2026 · PMID 42232895 · Full text
BACKGROUND AND OBJECTIVES: Elderly patients (≥65 years) who sustain burn injuries encounter a clinically significant perioperative challenge: a dysregulated hyperinflammatory response, characterized by elevated levels of...
BACKGROUND AND OBJECTIVES: Elderly patients (≥65 years) who sustain burn injuries encounter a clinically significant perioperative challenge: a dysregulated hyperinflammatory response, characterized by elevated levels of interleukin-6 (IL-6), tumor necrosis factor-alpha (TNF-α), and C-reactive protein (CRP), compounded by a markedly reduced hemodynamic reserve. Both propofol and low-dose ketamine exhibit distinct anti-inflammatory mechanisms; however, the optimization of their combined dosing within explicit safety parameters remains unestablished. Our objectives were to: (1) develop and externally validate a probabilistic machine learning (ML) model to predict dynamic 24-h trajectories of inflammatory markers; and (2) integrate these predictions with a safety-constrained offline reinforcement learning (RL) agent to formulate individualized propofol-ketamine dosing recommendations. STUDY DESIGN: This study employed a retrospective multi-cohort analysis utilizing two publicly accessible intensive care databases. SETTING: The research was conducted in an academic medical center ICU (MIMIC-IV) and across 208 community and academic hospitals (eICU Collaborative Research Database). MEASUREMENTS: The study analyzed 614 perioperative episodes in patients aged ≥65 years with confirmed burn injuries who received propofol-based anesthesia for ≥30 min and had ≥2 inflammatory laboratory measurements within 6-24 h post-induction. External validation was performed on 206 independent episodes. MAIN RESULTS: The proposed Event-Transformer with continuous-time Neural ODE dynamics demonstrated a 12-h IL-6 mean absolute error (MAE) of 6.82 pg/mL, representing a 70.1% improvement over linear mixed models (22.8 pg/mL). It achieved an inflammatory spike detection area under the receiver operating characteristic curve (AUROC) of 0.814 and empirical 90% prediction interval (PI) coverage of 87.2%. The Conservative Policy with Q-Learning (CPQL) dosing agent enhanced the time within the MAP target range (65-90 mmHg) from 62.3% to 71.8% ( < 0.001), decreased vasopressor initiation from 27.0% to 18.4% ( = 0.003), reduced peak predicted CRP by 21.3%, and decreased total propofol exposure by 12.1% through the introduction of adjunct ketamine (≈7.2 mcg/kg/min). The safety constraint violation rate was 0.0% under CPQL compared to 4.2% for unconstrained offline RL. CONCLUSIONS: An integrated inflammatory forecasting and dosing optimization pipeline can facilitate individualized propofol-ketamine titration in elderly burn patients, yielding predicted clinically significant improvements in hemodynamic stability and inflammatory burden, without safety violations. Clinically, the 70.1% reduction in IL-6 forecasting error translates to a meaningful difference between correct and incorrect inflammatory spike classification in a substantial fraction of patients, supporting the potential real-world utility of this framework as a decision-support tool to inform and guide future prospective trials.
Attention maps reveal stimulus-dependent retinal population codes.
Miqueles F, Palacios AG, Atkinson J … +1 more , Escobar MJ
Front Comput Neurosci · 2026 · PMID 42221576 · Full text
INTRODUCTION: Understanding how deep learning models map neural population activity to stimuli requires both high predictive accuracy and interpretable internal mechanisms. METHODS: In this work, we employ the POYO frame...
INTRODUCTION: Understanding how deep learning models map neural population activity to stimuli requires both high predictive accuracy and interpretable internal mechanisms. METHODS: In this work, we employ the POYO framework, a scalable transformer architecture based on spike tokenization and latent modeling, to decode large-scale retinal ganglion cell recordings. We ask whether the model's attention mechanisms can provide biologically meaningful insight by evaluating two contrasting conditions: a uniform flash stimulus and a spatiotemporally structured moving ball stimulus. RESULTS: We show that the model decodes both stimuli reliably and adapts rapidly to new preparations via fine-tuning, suggesting the capture of transferable population codes. We then analyze the model's internal organization, revealing that encoder attention patterns adapt to stimulus complexity: attention heads appear synchronized and broadly distributed for the flash stimulus, whereas they exhibit heterogeneous, specialized allocation strategies for the moving ball. By aggregating attention weights to identify the most relevant neurons for each task, we demonstrate that these high-attention units possess distinct physiological signatures-concentrating sustained, high-firing rates responses for the flash vs. diverse kinetics for the structured input. We confirm the causal validity of these findings via attention-guided ablations, where the progressive removal of these top-ranked units yields systematic losses in decoding performance. Furthermore, we expand the analysis to the decoder's attention, uncovering stimulus-specific retrieval strategies where individual heads exhibit distinct directional tuning preferences. DISCUSSION: We conclude that generic attention mechanisms can spontaneously recover biological coding strategies, identifying functionally distinct neural subpopulations without supervision, thus validating the utility of transformer-based architectures for neuroscientific discovery.
Temporal codes and recurrent timing nets for rhythmic expectancy.
Cariani P, Baker JM
Front Comput Neurosci · 2026 · PMID 42221575 · Full text
This paper focuses on possible time-domain neurocomputational mechanisms for short-term anticipatory processes. Here we present a simple, signal processing functional model of how short-term rhythmic pattern expectancies...
This paper focuses on possible time-domain neurocomputational mechanisms for short-term anticipatory processes. Here we present a simple, signal processing functional model of how short-term rhythmic pattern expectancies could be computed on the fly using recurrent neural timing nets (RTNs). The model is inspired by Gestaltist grouping principles for repeating temporal patterns of events (beats, pulses, grooves, metrical and non-metrical patterns). Building on previous autocorrelation models of pitch, meter, and rhythm, the RTN rhythm perception model consists of temporal codes, temporal pattern memory traces circulating in delay loops, and neural delay-and-coincidence networks with dynamically-adapting spike-correlation-dependent synapses. The network tracks in parallel all event periodicities in rhythmic hierarchies. As in memory trace theories of mismatch negativity (MMN-like) neural responses, it generates simple and complex pattern expectancies and registers deviations from them. Similarities and differences of this correlation-based model with those based on oscillators and predictive coding are discussed.
Coherent-resonant netting: disorder-enhanced selectivity from transient wave-like dynamics on biological connectomes.
Dolgikh O
Front Comput Neurosci · 2026 · PMID 42211247 · Full text
Biological agents face an energy-information bottleneck: inference requires rapid exploration of large hypothesis spaces, yet high-gain spiking is metabolically expensive. We propose Coherent-Resonant Netting (CRN) as a...
Biological agents face an energy-information bottleneck: inference requires rapid exploration of large hypothesis spaces, yet high-gain spiking is metabolically expensive. We propose Coherent-Resonant Netting (CRN) as a two-regime decision architecture in which a low-amplitude Stage-I transport process filters candidate routes on a structural graph before a higher-cost Stage-II commitment step. In this manuscript, we model Stage-I only, using a mechanistically neutral GKSL open-system proxy with dephasing rate κ and diagonal disorder ε. The model does not imply microscopic quantum coherence in neural tissue. In two biological connectome benchmarks, selectivity improves under partial coherence. In a compact touch-circuit benchmark, the wave proxy yields a 1.39 × improvement in peak target absorption over a matched low-temperature classical baseline. In a larva mushroom-body motif ( = 243 active nodes), selectivity shows a pronounced non-monotonic disorder-enhanced selectivity peak at intermediate ε, strongest in the native topology and strongly attenuated by degree-preserving rewiring. A permutation-based reanalysis confirms the pre-specified DES contrast (ε = 3 versus ε = 0, = 0.010), and the effect weakens progressively with increasing dephasing, becoming non-significant in the high-κ regime. We interpret these findings as evidence for a topology-sensitive, dephasing-dependent Stage-I routing effect on biological connectomes. Broader energetic and evolutionary implications remain conditional because Stage-II commitment is not explicitly modeled here.
Explainable hybrid CNN-transformer with self-supervised learning for structural analysis of paranasal sinus CT.
Ullah N, Algamdi SA, Sadad T
Front Comput Neurosci · 2026 · PMID 42211246 · Full text
INTRODUCTION: The process of precise structural evaluation for paranasal sinuses based on CT scan data establishes a foundation for medical professionals to assess human anatomical variations, supporting the diagnosis an...
INTRODUCTION: The process of precise structural evaluation for paranasal sinuses based on CT scan data establishes a foundation for medical professionals to assess human anatomical variations, supporting the diagnosis and treatment of ear, nose, and throat (ENT) conditions. Existing deep learning methods face difficulties in analyzing complex sinus structures due to limited annotated datasets and lack of clinical interpretability. METHODS: This study presents an explainable hybrid CNN-Transformer framework incorporating a self-supervised 3D convolutional autoencoder to perform structural analysis of paranasal sinus CT volumes. The framework is evaluated on the multi-institutional CT-SCOPE dataset, which contains diverse scan data from different hospitals and CT scanner models. The proposed approach combines anatomical segmentation to generate precise boundaries with residual-based structural representation learning for anomaly detection without requiring pathology labels. RESULTS: The hybrid segmentation model achieves high anatomical fidelity, producing Dice similarity coefficients above 0.83 across all four sinus regions, including maxillary, ethmoid, frontal, and sphenoid sinuses. The architecture integrates convolutional feature extraction with Transformer-based contextual modeling to capture both fine structural details and global anatomical context. The self-supervised autoencoder generates reconstruction residual maps that highlight structural deviations from standard osseous patterns. Grad-CAM-based interpretability analysis demonstrates that the model focuses on clinically relevant regions such as sinus walls and ethmoidal partitions. DISCUSSION: The integration of accurate segmentation, residual-based structural representation learning, and explainable attention mechanisms provides a comprehensive framework for anatomical analysis of paranasal sinus CT images. This approach establishes a foundation for future pathology-aware clinical modeling and supports the development of reliable AI-assisted diagnostic systems.
A predictive map learned from diverse entorhinal inputs explains the role of context-dependent reorganization of hippocampal place cells.
Kuniyoshi Y, Yamazaki T
Front Comput Neurosci · 2026 · PMID 42199560 · Full text
The hippocampus is thought to support spatial memory and navigation by constructing predictive representations of the environment. Predictive map theory formalizes this function as a successor representation (SR). Howeve...
The hippocampus is thought to support spatial memory and navigation by constructing predictive representations of the environment. Predictive map theory formalizes this function as a successor representation (SR). However, existing models assume a fixed and uniform distribution of place fields, despite experimental findings that place cell density is dynamically modulated by rewards and objects. Here, we propose a biologically inspired neural model in which predictive maps emerge from diverse entorhinal inputs. In the model, place cell-like representations are generated via non-negative sparse coding of medial entorhinal spatial signals and lateral entorhinal contextual and motivational signals, and are subsequently transformed into predictive maps using successor features. By coupling the predictive map to an actor-critic framework, the model supports goal-directed navigation in continuous environments. Furthermore, the model reproduces experience-dependent restructuring of hippocampal representations, including object-centered overrepresentation of place fields in two-dimensional environments and reward-centered overrepresentation in one-dimensional environments. Together, these results demonstrate that hippocampal predictive maps can emerge from the integration of diverse entorhinal inputs, providing a unified account of how spatial, contextual, and motivational information jointly shape hippocampal representations and behavior.
A brief history of dopamine prediction errors.
Dudhabhate BB, Costa KM
Front Comput Neurosci · 2026 · PMID 42179889 · Full text
Dopamine signaling has become closely associated with reward prediction errors (RPEs)-the difference between expected and experienced value. Although not without controversy, the dopamine RPE hypothesis is one of the mos...
Dopamine signaling has become closely associated with reward prediction errors (RPEs)-the difference between expected and experienced value. Although not without controversy, the dopamine RPE hypothesis is one of the most influential ideas in neuroscience. This review briefly summarizes its origins, empirical foundations, and theoretical development. We begin with early psychological studies which demonstrated that prediction errors, broadly defined, are central drivers of learning. These experiments inspired mathematical models that formalized associative learning rules and informed the development of reinforcement learning algorithms for artificial learning, including the influential temporal difference learning (TDRL) framework, where learning is guided by prediction errors in value or reward. These theoretical proposals converged with neuroscience through the landmark discovery that midbrain dopamine neurons show activity patterns that are strikingly similar to the RPEs proposed in TDRL. The idea that this unique neuronal population, already implicated in several behavioral processes and brain disorders, could encode a computational variable central to reinforcement learning algorithms was a major conceptual shift, and provided a strong framework that allowed for rigorous hypothesis testing. Over the past three decades, increasingly sophisticated experiments have both replicated the core dopamine RPE finding across distinct experimental contexts and revealed important deviations from the canonical model predictions. These exceptions have sparked ongoing debate about how the hypothesis should be enhanced, revised, or replaced. The history of the dopamine RPE hypothesis is a quintessential example of how the integration of theory and experiments can drive progress in neuroscience and offers a template for theoretical-experimental synthesis.
About
- Frequency
- Sun
- Papers found
- 200
- RSS feed
- Subscribe