Brain foundation models have achieved remarkable advances across a wide range of neuroscience tasks. However, most existing models are limited to a single functional modality, restricting their ability to exploit complementary spatiotemporal dynamics and the collective data scale across different neuroimaging techniques. This limitation largely arises from severe semantic heterogeneity and resolution discrepancies among modalities. To address these challenges, we propose Brain-OF, an omnifunctional brain foundation model jointly pretrained on fMRI, EEG and MEG, capable of handling both unimodal and multimodal inputs within a unified framework. To reconcile heterogeneous spatiotemporal resolutions, we introduce the Any-Resolution Neural Signal Sampler, which projects diverse brain signals into a shared semantic space. To further manage semantic shifts, the Brain-OF backbone integrates DINT attention with a Sparse Mixture of Experts, where shared experts capture modality-invariant representations and routed experts specialize in modality-specific semantics. Furthermore, to explicitly internalize the characteristics of neural activity through self-supervised learning, we propose Masked Temporal-Frequency Modeling, a dual-domain pretraining objective that jointly reconstructs brain signals in both the time and frequency domains. Brain-OF is pretrained on a large-scale corpus comprising around 40 datasets and demonstrates superior performance across diverse downstream tasks, highlighting the benefits of joint multimodal integration and dual-domain pretraining.

Diverse scientific and engineering research areas deal with discrete, time-stamped changes in large systems of interacting delay differential equations. Simulating such complex systems at scale on high-performance computing clusters demands efficient management of communication and memory. Inspired by the human cerebral cortex-a sparsely connected network of O(1010) neurons, each forming O(103)- O(104) synapses and communicating via short electrical pulses called spikes-we study the simulation of large-scale spiking neural networks for computational neuroscience research. This work presents a novel network construction method for multi-GPU clusters and upcoming exascale supercomputers using the message passing interface, where each process builds its local connectivity and prepares the data structures for efficient spike exchange across the cluster during state propagation. We demonstrate scaling performance of two cortical models using point-to-point and collective communication, respectively.

Parameter-Efficient Fine-Tuning (PEFT) methods have transformed the approach to fine-tuning large models for downstream tasks by enabling the adjustment of significantly fewer parameters than those in the original model matrices. In this work, we study the "very low rank regime", where we fine-tune the lowest amount of parameters per linear layer for each considered PEFT method. We propose 1LoRA (Summation Low-Rank Adaptation), a compute, parameter and memory efficient fine-tuning method which uses the feature sum as fixed compression and a single trainable vector as decompression. Differently from state-of-the-art PEFT methods like LoRA, VeRA, and the recent MoRA, 1LoRA uses fewer parameters per layer, reducing the memory footprint and the computational cost. We extensively evaluate our method against state-of-the-art PEFT methods on multiple fine-tuning tasks, and show that our method not only outperforms them, but is also more parameter, memory and computationally efficient. Moreover, thanks to its memory efficiency, 1LoRA allows to fine-tune more evenly across layers, instead of focusing on specific ones (e.g. attention layers), improving performance further.

Spiking Neural Networks (SNNs) have gained significant attention in recent years because their spike-based, event-driven communication, together with the inherent sparsity of spikes, is believed to enable highly energy-efficient computation. This property has been effectively leveraged in a variety of classification and regression tasks. Although inspired by biological neural systems, most existing SNNs bear only limited resemblance to their biological counterparts in terms of dynamics and learning mechanisms. Several studies have aimed to better understand the mathematical functioning of the brain to incorporate greater biological plausibility into existing neuron models. One such approach is to train the networks using a more biologically plausible learning rule called eligibility propagation (e-prop), which serves as an alternative to conventional Back Propagation Through Time (BPTT). In this study, we train a recurrent-SNN composed of Adaptive-threshold based Leaky Integrate-and-Fire (ALIF) neurons using the e-prop learning rule on datasets derived from viscoelastic rheological models. We establish an analogy between these rheological models and the membrane dynamics of LIF neurons, highlighting their mathematical similarity and providing a physical justification for employing SNNs in mechanics beyond energy-efficiency considerations. To model geometrically and physically nonlinear structural deformations, we further propose a multi-layer recurrent spiking architecture that integrates Legendre Memory Units with ALIF neurons for predicting nonlinear transient displacement responses of plates obtained from shock tube experiments. To train this architecture, a hybrid learning approach is introduced by combining e-prop and BPTT. Finally, a comparative analysis between e-prop and BPTT is conducted, emphasizing their respective trade-offs in accuracy, convergence speed, and computational efficiency.

Self-supervised learning (SSL) methods have achieved remarkable success in learning image representations allowing invariances in them — but therefore discarding transformation information that some computer vision tasks actually require. While recent approaches attempt to address this limitation by learning equivariant features using linear operators in feature space, they impose restrictive assumptions that constrain flexibility and generalization. We introduce a weaker definition for the transformation relation between image and feature space denoted as equivariance-coherence. We propose a novel SSL auxillary task that learns equivariance-coherent representations through intermediate transformation reconstruction, which can be integrated with existing joint embedding SSL methods. Our key idea is to reconstruct images at intermediate points along transformation paths, e.g. when training on 30° rotations, we reconstruct the 10° and 20° rotation states. Reconstructing intermediate states requires the transformation information used in augmentations, rather than suppressing it, and therefore fosters features containing the augmented transformation information. Our method decomposes feature vectors into invariant and equivariant parts, training them with standard SSL losses and reconstruction losses, respectively. We demonstrate substantial improvements on synthetic equivariance benchmarks while maintaining competitive performance on downstream tasks requiring invariant representations. The approach seamlessly integrates with existing SSL methods (iBOT, DINOv2) and consistently enhances performance across diverse tasks, including segmentation, detection, depth estimation, and video dense prediction. Our framework provides a practical way for augmenting SSL methods with equivariant capabilities while preserving invariant performance.

While the implementation of learning and memory in the brain is governed in large part by subcellular mechanims in the dendrites of neurons, large-scale network simulations featuring such processes remain challenging to achieve. This can be attributed to a lack of appropriate software tools, as neuroscientific simulation software focuses on the one hand on highly detailed models, and on the other hand on massive networks featuring point-neurons. Here, we fill this gap by implementing a framework for the massively parallel simulation of simplified dendrite models with customizable subcellular dynamics. To achieve this, we leverage the NEural Simulation Tool (NEST), the neuroscientific reference with respect to efficient massively parallel simulations of point neuron networks. By co-opting the already existing model descriptions language NESTML, we generate C++ code implementing user-configurable subcellular dynamics. Through benchmarking and profiling, we show that the generated models run efficiently, leading to scalable NEST network simulations. We demonstrate relevant functionalities by showing that a key sensory computation – the association of top-down context arriving at distal dendrites in layer 1 and feedforward sensory input arriving perisomatically – can be achieved in a single shot fashion through apical calcium dynamics. Our work thus unlocks the study of how dendritic processes shape learning, and in particular of how brain-wide communication through long-range, layer 1-targetting connections steers perisomatic plasticity. ### Competing Interest Statement The authors have declared no competing interest. Helmholtz Association of German Research Centres, https://ror.org/0281dp749 Italian Ministry for Universities and Research (MUR), CUP I53C22001400006, CUP B51E22000150006

Monocular depth estimation (MDE) is a challenging task in computer vision, often hindered by the cost and scarcity of high-quality labeled datasets. We tackle this challenge using auxiliary datasets from related vision tasks for an alternating training scheme with a shared decoder built on top of a pre-trained vision foundation model, while giving a higher weight to MDE. Through extensive experiments we demonstrate the benefits of incorporating various in-domain auxiliary datasets and tasks to improve MDE quality on average by ~ 11 %. Our experimental analysis shows that auxiliary tasks have different impacts, confirming the importance of task selection, highlighting that quality gains are not achieved by merely adding data. Remarkably, our study reveals that using semantic segmentation datasets as Multi-Label Dense Classification (MLDC) often results in additional quality gains. Lastly, our method significantly improves the data efficiency for the considered MDE datasets, enhancing their quality while reducing their size by at least 80%. This paves the way for using auxiliary data from related tasks to improve MDE quality despite limited availability of high-quality labeled data. Code is available at https://jugit.fz-juelich.de/ias-8/mdeaux.

In deep learning, using larger training datasets usually leads to more accurate models. However, simply adding more but redundant data may be inefficient, as some training samples may be more informative than others. We propose Focal Sampling, a method that biases SGD (Stochastic Gradient Descent) towards samples that are found to be more important after a few training epochs, by sampling them more often for the rest of the training. In contrast to state-of-the-art, our approach requires less computational overhead to estimate sample importance, as it computes estimates once during training using the prediction probabilities, and does not require restarting training. In the experimental evaluation, we see that our learning technique trains faster than state-of-the-art and can achieve higher test accuracy, especially when datasets are not well balanced or when using multiple data augmentations. Lastly, results suggest that our approach has intrinsic balancing properties and that balancing datasets based on class importance, rather than by number of samples, can achieve higher test accuracy. Code is available at https://jugit.fz-juelich.de/ias-8/sgd_biased.

With increasing model complexity, models are typically re-used and evolved rather than starting from scratch. There is also a growing challenge in ensuring that these models can seamlessly work across various simulation backends and hardware platforms. This underscores the need to ensure that models are easily findable, accessible, interoperable, and reusable—adhering to the FAIR principles. NESTML addresses these requirements by providing a domain-specific language for describing neuron and synapse models that covers a wide range of neuroscientific use cases. The language is supported by a code generation toolchain that automatically generates low-level simulation code for a given target platform (for example, C++ code targeting NEST Simulator). Code generation allows an accessible and easy-to-use language syntax to be combined with good runtime simulation performance and scalability. With an intuitive and highly generic language, combined with the generation of efficient, optimized simulation code supporting large-scale simulations, it opens up neuronal network model development and simulation as a research tool to a much wider community. While originally developed in the context of NEST Simulator, NESTML has been extended to target other simulation platforms, such as the SpiNNaker neuromorphic hardware platform. The processing toolchain is written in Python and is lightweight and easily customizable, making it easy to add support for new simulation platforms.

Brain vessel segmentation of MR scans is a critical step in the diagnosis of cerebrovascular diseases. Due to the fine vessel structure, manual vessel segmentation is time consuming. Therefore, automatic deep learning (DL) based segmentation techniques are intensively investigated. As conventional DL models yield a high complexity and lack an indication of decision reliability, they are often considered as not trustworthy. This work aims to increase trust in DL based models by incorporating epistemic uncertainty quantification into cerebrovascular segmentation models for the first time. By implementing an efficient ensemble model combining the advantages of Bayesian Approximation and Deep Ensembles, we aim to overcome the high computational costs of conventional probabilistic networks. Areas of high model uncertainty and erroneous predictions are aligned which demonstrates the effectiveness and reliability of the approach. We perform extensive experiments applying the ensemble model on out-of-distribution (OOD) data. We demonstrate that for OOD-images, the estimated uncertainty increases. Additionally, omitting highly uncertain areas improves the segmentation quality, both for in- and out-of-distribution data. The ensemble model explains its limitations in a reliable manner and can maintain trustworthiness also for OOD data and could be considered in clinical applications

Sequential structure is a key feature of multiple domains of natural cognition and behavior, such as language, movement and decision-making. Likewise, it is also a central property of tasks to which we would like to apply artificial intelligence. It is therefore of great importance to develop frameworks that allow us to evaluate sequence learning and processing in a domain agnostic fashion, whilst simultaneously providing a link to formal theories of computation and computability. To address this need, we introduce two complementary software tools: SymSeq, designed to rigorously generate and analyze structured symbolic sequences, and SeqBench, a comprehensive benchmark suite of rule-based sequence processing tasks to evaluate the performance of artificial learning systems in cognitively relevant domains. In combination, SymSeqBench offers versatility in investigating sequential structure across diverse knowledge domains, including experimental psycholinguistics, cognitive psychology, behavioral analysis, neuromorphic computing and artificial intelligence. Due to its basis in Formal Language Theory (FLT), SymSeqBench provides researchers in multiple domains with a convenient and practical way to apply the concepts of FLT to conceptualize and standardize their experiments, thus advancing our understanding of cognition and behavior through shared computational frameworks and formalisms. The tool is modular, openly available and accessible to the research community.

Understanding information propagation and computational capabilities in large-scale brain models is crucial for advancing both neuroscience and neuro-inspired computing. It remains unknown how complexity affects the performance of simulated biological networks in predicting non-linear dynamics. Here, this issue is addressed by integrating reservoir computing, a recurrent neural network architecture, with The Virtual Brain, a whole-brain simulation platform, to create a more neurophysiologically-plausible machine learning framework. Metrics derived from nonlinear dynamics and complexity theory, including the largest Lyapunov exponent, which captures chaotic behavior, Detrended Fluctuation Analysis, which assesses temporal correlations, and the Perturbational Complexity Index, which evaluates the complexity of neural responses, provide a quantitative framework for characterizing cognitive dynamics. Deploying this framework on High Performance Computing enables a thorough exploration of the vast parameter space, utilizing a diverse evaluation framework that assesses simulations through these metrics, implemented on Graphics Processing Units (GPUs). This enables the identification of optimal parameter regimes, a comprehensive characterization of the complex dynamics exhibited by the system, and a deeper understanding of the underlying mechanisms governing the TVB-based reservoir's computational capabilities. Ridge regression, accelerated also by GPUs, is used to extract the predictive capacity from the reservoir states. The results suggest that edge-of-chaos dynamics correspond to enhanced memory and prediction accuracy, supporting the potential of TVB-based reservoirs for brain-inspired machine learning.

Integration of information across heterogeneous sources creates added scientific value. Interoperability of data, tools and models is, however, difficult to accomplish across spatial and temporal scales. Here we introduce the toolbox Parallel Co-Simulation, which enables the interoperation of simulators operating at different scales. We provide a software science co-design pattern and illustrate its functioning along a neuroscience example, in which individual regions of interest are simulated on the cellular level allowing us to study detailed mechanisms, while the remaining network is efficiently simulated on the population level. A workflow is illustrated for the use case of The Virtual Brain and NEST, in which the CA1 region of the cellular-level hippocampus of the mouse is embedded into a full brain network involving micro and macro electrode recordings. This new tool allows integrating knowledge across scales in the same simulation framework and validating them against multiscale experiments, thereby largely widening the explanatory power of computational models.

In deep learning, using larger training datasets usually leads to more accurate models. However, simply adding more but redundant data may be inefficient, as some training samples may be more informative than others. We propose to bias SGD (Stochastic Gradient Descent) towards samples that are found to be more important after a few training epochs, by sampling them more often for the rest of training.In contrast to state-of-the-art, our approach requires less computational overhead to estimate sample importance, as it computes estimates once during training using the prediction probabilities, and does not require that training be restarted.In the experimental evaluation, we see that our learning technique trains faster than state-of-the-art and can achieve higher test accuracy, especially when datasets are not well balanced. Lastly, results suggest that our approach has intrinsic balancing properties. Code is available at https://github.com/AlessioQuercia/sgd_biased.

To acquire statistical regularities from the world, the brain must reliably process, and learn from, spatio-temporally structured information. Although an increasing number of computational models have attempted to explain how such sequence learning may be implemented in the neural hardware, many remain limited in functionality or lack biophysical plausibility. If we are to harvest the knowledge within these models and arrive at a deeper mechanistic understanding of sequential processing in cortical circuits, it is critical that the models and their findings are accessible, reproducible, and quantitatively comparable. Here we illustrate the importance of these aspects by providing a thorough investigation of a recently proposed sequence learning model. We re-implement the modular columnar architecture and reward-based learning rule in the open-source NEST simulator, and successfully replicate the main findings of the original study. Building on these, we perform an in-depth analysis of the model's robustness to parameter settings and underlying assumptions, highlighting its strengths and weaknesses. We demonstrate a limitation of the model consisting in the hard-wiring of the sequence order in the connectivity patterns, and suggest possible solutions. Finally, we show that the core functionality of the model is retained under more biologically-plausible constraints.

While sensory representations in the brain depend on context, it remains unclear how such modulations are implemented at the biophysical level, and how processing layers further in the hierarchy can extract useful features for each possible contextual state. Here, we demonstrate that dendritic N-Methyl-D-Aspartate spikes can, within physiological constraints, implement contextual modulation of feedforward processing. Such neuron-specific modulations exploit prior knowledge, encoded in stable feedforward weights, to achieve transfer learning across contexts. In a network of biophysically realistic neuron models with context-independent feedforward weights, we show that modulatory inputs to dendritic branches can solve linearly nonseparable learning problems with a Hebbian, error-modulated learning rule. We also demonstrate that local prediction of whether representations originate either from different inputs, or from different contextual modulations of the same input, results in representation learning of hierarchical feedforward weights across processing layers that accommodate a multitude of contexts.

Simulation speed matters for neuroscientific research: this includes not only how quickly the simulated model time of a large-scale spiking neuronal network progresses, but also how long it takes to instantiate the network model in computer memory. On the hardware side, acceleration via highly parallel GPUs is being increasingly utilized. On the software side, code generation approaches ensure highly optimized code, at the expense of repeated code regeneration and recompilation after modifications to the network model. Aiming for a greater flexibility with respect to iterative model changes, here we propose a new method for creating network connections interactively, dynamically, and directly in GPU memory through a set of commonly used high-level connection rules. We validate the simulation performance with both consumer and data center GPUs on two neuroscientifically relevant models: a cortical microcircuit of about 77,000 leaky-integrate-and-fire neuron models and 300 million static synapses, and a two-population network recurrently connected using a variety of connection rules. With our proposed ad hoc network instantiation, both network construction and simulation times are comparable or shorter than those obtained with other state-of-the-art simulation technologies, while still meeting the flexibility demands of explorative network modeling.

Information from the sensory periphery is conveyed to the cortex via structured projection pathways that spatially segregate stimulus features, providing a robust and efficient encoding strategy. Beyond sensory encoding, this prominent anatomical feature extends throughout the neocortex. However, the extent to which it influences cortical processing is unclear. In this study, we combine cortical circuit modeling with network theory to demonstrate that the sharpness of topographic projections acts as a bifurcation parameter, controlling the macroscopic dynamics and representational precision across a modular network. By shifting the balance of excitation and inhibition, topographic modularity gradually increases task performance and improves the signal-to-noise ratio across the system. We demonstrate that in biologically constrained networks, such a denoising behavior is contingent on recurrent inhibition. We show that this is a robust and generic structural feature that enables a broad range of behaviorally relevant operating regimes, and provide an in-depth theoretical analysis unraveling the dynamical principles underlying the mechanism.