Gamma synchrony is ubiquitous in visual cortex, but whether it contributes to perceptual grouping remains contentious based on observations that gamma frequency is not consistent across stimulus features and that gamma synchrony depends on distances between image elements. These stimulus dependencies have been argued to challenge the idea that the visual system groups image elements by synchronizing the neural assemblies that encode them. Here, we argue instead that these dependencies may shape synchrony in perceptually meaningful ways. Indeed, according to the theory of weakly coupled oscillators (TWCO), synchrony-based grouping mechanisms require stimulus dependence. Synchronization among coupled oscillators depends on frequency dissimilarity and coupling strength, which in early visual cortex relate to local feature dissimilarity and physical distance, respectively. We manipulated these factors in a texture segregation experiment wherein human observers identified the orientation of a figure defined by reduced contrast heterogeneity compared to the background. Human performance followed TWCO predictions both qualitatively and quantitatively, as formalized in a computational model. Moreover, we found that when enriched with a Hebbian learning rule, our model also predicted human learning effects: Increases in model gamma synchrony due to perceptual learning predicted improvements in texture segregation across sessions. Taken together, our data suggest that the stimulus-dependence of gamma synchrony captures local image statistics and is linked to the stimulus-dependence of texture segregation, and that the effect of visual experience on gamma synchrony provides a viable perceptual learning mechanism for training-induced improvements in texture segregation. Our results suggest that gamma synchrony with its inherent stimulus dependencies can provide a plausible mechanistic basis for perceptual grouping and visual scene segmentation.

Contemporary computational neuroscience features two prominent modeling traditions. Bottom-up whole-brain modeling (WBM) builds biophysically detailed simulations of brain structure and dynamics, whereas top-down neuroconnectionism optimizes deep neural networks for functional performance. Each has achieved remarkable success yet remains incomplete with WBMs lacking functional competence and neuroconnectionist models showing limited biological grounding. Here we propose functional whole-brain models (fWBMs) as a unified modeling paradigm that integrates structural and dynamical realism with task-performing capacity. fWBMs are defined by four minimal criteria: structural grounding in empirical connectomes and regional biology, continuous-time dynamical realism, functional competence across cognitive domains, and mappable observables to neuroimaging, electrophysiologcal and behavioral data. To formalize this integration, we establish a three-pillar roadmap across short-, mid-, and long-term horizons, and outline the scientific and clinical opportunities this paradigm enables. We argue that the disciplined pursuit of this integrative vision will generate the tools, common language, and cross-scale hypotheses needed to advance our understanding of the brain.

Neuroscience is rapidly growing, rendering it hard to keep track of the field’s structural organization, research trends, and open questions. This study tackles this issue by analyzing a vast corpus of articles published between 1999 and 2023. Findings reveal a field with a strong experimental focus, encompassing both hypothesis-driven and data-driven methodologies and an increasing use of advanced data analytics and artificial intelligence. A growing emphasis on applied research is evident, particularly in areas like neurodegeneration, neuromodulation, and technological advancements, while fundamental research is at risk of decline. The field demonstrates high levels of interdisciplinarity and surprisingly robust cross-fertilization across research domains, with key intellectual hubs shaping the broader landscape. Nevertheless, neuroscience widely relies on specific mechanistic explanations rather than unifying theoretical frameworks, and the integration of findings across different spatial and temporal scales remains limited. This study provides a framework for understanding neuroscience’s trajectory and identifies potential avenues for strengthening the field.

Forecasting neural activity in response to naturalistic stimuli remains a key challenge for understanding brain dynamics and enabling downstream neurotechnological applications. Here, we introduce a generative forecasting framework for modeling neural dynamics based on autoregressive flow matching (AFM). Building on recent advances in transport-based generative modeling, our approach probabilistically predicts neural responses at scale from multimodal sensory input. Specifically, we learn the conditional distribution of future neural activity given past neural dynamics and concurrent sensory input, explicitly modeling neural activity as a temporally evolving process in which future states depend on recent neural history. We evaluate our framework on the Algonauts project 2025 challenge functional magnetic resonance imaging dataset using subject-specific models. AFM significantly outperforms both a non-autoregressive flow-matching baseline and the official challenge general linear model baseline in predicting short-term parcel-wise blood oxygenation level-dependent (BOLD) activity, demonstrating improved generalization and widespread cortical prediction performance. Ablation analyses show that access to past BOLD dynamics is a dominant driver of performance, while autoregressive factorization yields consistent, modest gains under short-horizon, context-rich conditions. Together, these findings position autoregressive flow-based generative modeling as an effective approach for short-term probabilistic forecasting of neural dynamics with promising applications in closed-loop neurotechnology.

Bistable perception involves the spontaneous alternation between two exclusive interpretations of a single stimulus. Previous research has suggested that this perceptual phenomenon results from winnerless dynamics in the cortex. Indeed, winnerless dynamics can explain many key behavioral characteristics of bistable perception. However, it fails to explain an increase in alternation rate that is typically observed in response to increased stimulus drive and instead predicts a decline in alternation rate. To reconcile this discrepancy, several lines of work have augmented winnerless dynamics with additional processes such as global gain control, input suppression, and release mechanisms. These offer potential explanations at an algorithmic level. But it remains unclear which, if any, of these mechanisms are implemented in the cortex and what their biological substrates might be. We suggest that the answers to these questions may lie within the laminar architecture of the cortical microcircuit. Utilizing a dynamic mean field approach, we implement a laminar columnar circuit with empirically derived interlaminar connectivity. By coupling two such circuits such that they exhibit competition, we are able to produce winnerless dynamics reflective of bistable perception. Within our model, we show that two mechanisms emerge from the layered structure that ensure increased alternation rate in response to increased stimulus drive. First, deep layers act to inhibit the upper layers, thereby reducing the attractor depth and increasing the alternation rate. Second, recurrent connections between superficial and granular layers implement an input suppression mechanism which again reduces the attractor depth of the winnerless competition. These findings suggest the functional significance of the layered cortical architecture as they showcase perceptual implications of neuroanatomical properties such as interlaminar connectivity and layer-specific activation.

Hierarchical predictive coding (hPC) provides a compelling framework for understanding how the cortex predicts future sensory inputs by minimizing prediction errors through an internal generative model of the external world. Existing deep learning models inspired by hPC incorporate architectural choices that deviate from core hPC principles, potentially limiting their utility for neuroscientific investigations. We introduce PrediRep (Predicting Representations), a novel deep learning network that adheres more closely to architectural principles of hPC. We validate PrediRep by comparing its functional alignment with hPC to that of existing models after being trained on a next-frame prediction task. Our findings demonstrate that PrediRep, particularly when trained with an all-level loss function (PrediRepAll), exhibits high functional alignment with hPC. In contrast to other contemporary deep learning networks inspired by hPC, it consistently processes input-relevant information at higher hierarchical levels and maintains active representations and accurate predictions across all hierarchical levels. Although PrediRep was designed primarily to serve as a model suitable for neuroscientific research rather than to optimize performance, it nevertheless achieves competitive performance in next-frame prediction while utilizing significantly fewer trainable parameters than alternative models. Our results underscore that even minor architectural deviations from neuroscientific theories like hPC can lead to significant functional discrepancies. By faithfully adhering to hPC principles, PrediRep provides a more accurate tool for in silico exploration of cortical phenomena. PrediRep’s lightweight and biologically plausible design makes it well-suited for future studies aiming to investigate the neural underpinnings of predictive coding and to derive empirically testable predictions.

Neuroscience emerged as a distinct academic discipline during the 20th century and has undergone rapid expansion since then. However, a comprehensive analysis of its evolving landscape remains lacking. The present study provides a large-scale, data-driven, mapping of neuroscience research. Leveraging large language models and clustering techniques to analyze 461,316 articles published between 1999 and 2023, this study reveals the field's structural organization and highlights dominant themes. Citation network analysis uncovers a surprisingly integrated picture and key intellectual hubs that shape the broader research landscape. An analysis of how research clusters align with pre-defined dimensions demonstrates a strong experimental focus, widespread reliance on micro theories, and a growing emphasis on applied and translational research. At the same time, fundamental research is at the risk of decline while theoretical work and cross-scale integration remain limited. This study provides a framework for understanding neuroscience's trajectory and identifies potential avenues for strengthening the field. ### Competing Interest Statement The authors have declared no competing interest.

Understanding how cortex, subcortex and cerebellum integrate is a major challenge for neuroscience, however, studies of the brain's structural connectivity have mostly focused on cortico-cortical links. Here, we used diffusion imaging to construct the structural connectome of the entire human brain including 360 cortical, 233 subcortical, and 125 cerebellar regions of interest (ROIs). We found that the brain forms a modular and hierarchical network architecture, organized into modules of mixed cortical, subcortical and/or cerebellar regions, and whose cross-modular pathways are centralized through highly connected hub ROIs (a `rich-club'). This global rich-club is subcortically dominated and, surprisingly, composed of hub ROIs from all subcortical structures rather than one region like the thalamus, centralizing the communication pathways. This study improves our understanding of the human brain's organization. It provides structural evidence to question the prevalent cortico-centric notion by revealing a connectome centered at the subcortex but made of transversal pathways.

This Perspective presents the Modular-Integrative Modeling approach, a novel framework in neuroscience for developing brain models that blend biological realism with functional performance to provide a holistic view on brain function in interaction with the body and environment.

Although the structure of cortical networks provides the necessary substrate for their neuronal activity, the structure alone does not suffice to understand the activity. Leveraging the increasing availability of human data, we developed a multi-scale, spiking network model of human cortex to investigate the relationship between structure and dynamics. In this model, each area in one hemisphere of the Desikan-Killiany parcellation is represented by a $1\,\mathrm{mm^{2}}$ column with a layered structure. The model aggregates data across multiple modalities, including electron microscopy, electrophysiology, morphological reconstructions, and diffusion tensor imaging, into a coherent framework. It predicts activity on all scales from the single-neuron spiking activity to the area-level functional connectivity. We compared the model activity with human electrophysiological data and human resting-state functional magnetic resonance imaging (fMRI) data. This comparison reveals that the model can reproduce aspects of both spiking statistics and fMRI correlations if the inter-areal connections are sufficiently strong. Furthermore, we study the propagation of a single-spike perturbation and macroscopic fluctuations through the network. The open-source model serves as an integrative platform for further refinements and future in silico studies of human cortical structure, dynamics, and function.

Primate visual cortex exhibits key organizational principles: cortical magnification, eccentricity-dependent receptive field size and spatial frequency tuning as well as radial bias. We provide compelling evidence that these principles arise from the interplay of the non-uniform distribution of retinal ganglion cells, and a quasi-uniform convergence rate from the retina to the cortex. We show that convolutional neural networks outfitted with a retinal sampling layer, which resamples images according to retinal ganglion cell density, develop these organizational principles. Surprisingly, our results indicate that radial bias is spatial-frequency dependent and only manifests for high spatial frequencies. For low spatial frequencies, the bias shifts towards orthogonal orientations. These findings introduce a novel hypothesis about the origin of radial bias. Quasi-uniform convergence limits the range of spatial frequencies (in retinal space) that can be resolved, while retinal sampling determines the spatial frequency content throughout the retina.

Primate visual cortex exhibits key organizational principles: Cortical magnification, eccentricity-dependent receptive field size and spatial frequency tuning as well as radial bias. We provide compelling evidence that these principles arise from the interplay of the non-uniform distribution of retinal ganglion cells (RGCs), and a quasi-uniform convergence rate from the retina to the cortex. We show that convolutional neural networks (CNNs) outfitted with a retinal sampling layer, which resamples images according to retinal ganglion cell density, develop these organizational principles. Surprisingly, our results indicate that radial bias is spatial-frequency dependent and only manifests for high spatial frequencies. For low spatial frequencies, the bias shifts towards orthogonal orientations. These findings introduce a novel hypothesis about the origin of radial bias. Quasi-uniform convergence limits the range of spatial frequencies (in retinal space) that can be resolved, while retinal sampling determines the spatial frequency content throughout the retina.

ABSTRACT The white matter is made of anatomical fibres that constitute the highway of long-range connections between different parts of the brain. This network is referred to as the brain’s structural connectivity and lays the foundation of network interaction between brain areas. When analysing the architectural principles of this global network most studies have mainly focused on cortico-cortical and partly on cortico-subcortical connections. Here we show, for the first time, how the integrated cortical, subcortical, and cerebellar brain areas shape the structural architecture of the whole brain. We find that dense clusters vertically transverse cortical, subcortical, and cerebellar brain areas, which are themselves centralised by a global rich-club consisting similarly of cortical and subcortical brain areas. Notably, the most prominent hubs can be found in subcortical brain regions, and their targeted in-silico lesions proved to be most harmful for global signal propagation. Individually, the cortical, subcortical, and cerebellar sub-networks manifest distinct network features despite some similarities, which underline their unique structural fingerprints. Our results, exposing the heterogeneity of internal organisation across cortex, subcortex, and cerebellum, and the crucial role of the subcortex for the integration of the global anatomical pathways, highlight the need to overcome the prevalent cortex-centric focus towards a global consideration of the structural connectivity.

Goal-driven deep learning increasingly supplements classical modeling approaches in computational neuroscience. The strength of deep neural networks as models of the brain lies in their ability to autonomously learn the connectivity required to solve complex and ecologically valid tasks, obviating the need for hand-engineered or hypothesis-driven connectivity patterns. Consequently, goal-driven models can generate hypotheses about the neurocomputations underlying cortical processing that are grounded in macro- and mesoscopic anatomical properties of the network's biological counterpart. Whereas, goal-driven modeling is already becoming prevalent in the neuroscience of perception, its application to the sensorimotor domain is currently hampered by the complexity of the methods required to train models comprising the closed sensation-action loop. This paper describes AngoraPy, a Python library that mitigates this obstacle by providing researchers with the tools necessary to train complex recurrent convolutional neural networks that model the human sensorimotor system. To make the technical details of this toolkit more approachable, an illustrative example that trains a recurrent toy model on in-hand object manipulation accompanies the theoretical remarks. An extensive benchmark on various classical, 3D robotic, and anthropomorphic control tasks demonstrates AngoraPy's general applicability to a wide range of tasks. Together with its ability to adaptively handle custom architectures, the flexibility of this toolkit demonstrates its power for goal-driven sensorimotor modeling.