Neurocomputing

A number of potentially important mechanisms have been identified as key players to generate epileptiform activity, such as genetic mutations, activity-dependent alteration of synaptic functions, and functional network reorganization at the macroscopic level. Here we study how network connectivity at cellular level can affect the onset of epileptiform activity, using computational model networks with different wiring properties. The model suggests that networks connected as in real brain circuits are more resistant to generate seizure-like activity. The results suggest new experimentally testable predictions on the cellular network connectivity in epileptic individuals, and highlight the importance of using the appropriate network connectivity to investigate epileptiform activity with computational models.

This research report introduces ElegansNet, a neural network that mimics real-world neuronal network circuitry, with the goal of better understanding the interplay between connectome topology and deep learning systems. The proposed approach utilizes the powerful representational capabilities of living beings neuronal circuitry to design and generate improved deep learning systems with a topology similar to natural networks. The Caenorhabditis elegans connectome is used as a reference due to its completeness, reasonable size, and functional neuron classes annotations. It is demonstrated that the connectome of simple organisms exhibits specific functional relationships between neurons, and once transformed into learnable tensor networks and integrated into modern architectures, it offers bio-plausible structures that efficiently solve complex tasks. The performance of the models is demonstrated against randomly wired networks and compared to artificial networks ranked on global benchmarks. In the first case, ElegansNet outperforms randomly wired networks. Interestingly, ElegansNet models show slightly similar performance with only those based on the Watts-Strogatz small-world property. When compared to state-of-the-art artificial neural networks, such as transformers or attention-based autoencoders, ElegansNet outperforms well-known deep learning and traditional models in both supervised image classification tasks and unsupervised hand-written digits reconstruction, achieving top-1 accuracy of 99.99% on Cifar10 and 99.84% on MNIST Unsup on the validation sets.

In the brain, many regions work in a network-like association, yet it is not known how durable these associations are in terms of activity and could survive without structural connections. To assess the association or similarity between brain regions with a new "generating" approach, this study evaluated the similarity of activities of neurons at the cellular level within each region after disconnecting between regions. To this end, a multi-layer LSTM (Long-Short Term Memory) model was used. Surprisingly, the results revealed that generation of activity from one region to other regions that had been disconnected was possible with similar reproduction accuracy as generation between the same regions in many cases. Notably, not only firing rates but also synchronization of firing between neuron pairs, which is often used as neuronal representations, could be reproduced with considerable precision. Additionally, their accuracies were associated with the relative distance between brain regions and the strength of the structural connections that initially connected them. This outcome not only enables us to look into principles in neuroscience based on the potential to generate new informative data, but also creates neural activity that has not been measured in adequate amounts and could potentially lead to reduced animal experiments.

Deep neural networks applied to signal processing tasks often need specialized architectural mechanisms to capture the temporal history of input signals. Traditional approaches include recurrent loops between layers, gated units, or tapped delay lines. However, biological brains exhibit much richer dynamics, characterized by activity across multiple frequency bands (alpha, beta, gamma, delta) and phenomena such as phase locking and synchronization. Standard Recurrent Neural Networks (RNNs) are limited in their ability to represent these complex dynamical features. In this work, we introduce a novel framework called the Deep Coupled Kuramoto Oscillatory Neural Network (DcKONN), which leverages networks of nonlinear Kuramoto oscillators trained in a deep learning paradigm. The DcKONN architecture has been applied to EEG signal classifier task. Simulation results demonstrate that the proposed oscillatory neural networks achieve superior or comparable classification accuracy compared to existing state-of-the-art models. Beyond performance improvements, these models also provide valuable neurobiological insights by naturally incorporating oscillatory dynamics into their architecture.

The brain is believed to process information efficiently in a different manner from deep learning-based artificial intelligence (AI). Brain-like next-generation AI is gaining attention owing to its potential to perform human-like, highly adaptive, robust, and power-efficient computation. To realize such AI, one crucial approach is the bottom-up implementation of the neuronal systems, capturing their electrophysiological characteristics in electronic circuits. However, this neuromorphic approach generally focuses on simplified neuronal models that do not refer to many biological findings. Developing closer-to-brain models is a natural direction that serve as a fundamental computing model for next-generation AI. One of the constraints of neuromorphic circuits is the bit resolution of synaptic efficacy memory, as the memory footprint scales with it precision. Although low-resolution synaptic efficacy is essential for minimizing memory circuit footprint and energy consumption, it generally leads to performance degradation in many tasks such as the spatio-temporal spike pattern detection. This study proposed a closer-to-brain learning rule that incorporates heterosynaptic plasticity (HP) induced by glutamate spillover. It is demonstrated that our model mitigates the performance degradation associated with low-bit resolution synaptic efficacy, achieving the pattern detection success rate with 3-bit resolution synaptic efficacy, which is comparable to 64-bit floating-point precision. Furthermore, the findings of the study indicate that HP based model accelerates the convergence of the synaptic effcacy and effectively potentiates the synapses relevant to the pattern detection while suppressing irrelevant ones, thereby promoting a bimodal distribution of synaptic efficacies. These findings may provide a basic framework for constructing an energy-efficient, brain-like next-generation AI that maintains high performance under hardware constraints.

In this study, we investigate the use of temporal dynamics in brain connectivity for the classification of electroencephalography (EEG) signals using dynamic Graph Neural Networks (GNNs). Our methods are applied to several large-scale EEG datasets focused on abnormality and epilepsy detection. The implemented models demonstrate competitive performance on unseen test subjects across all three datasets, outperforming previous graph-based baselines in terms of accuracy and F1 score. We explore multiple architectures designed to capture temporal variations in graph-structured data, demonstrating their effectiveness in modeling dynamic brain activity. In addition to classification, we employ graph-theoretical metrics to analyze temporal changes in brain networks, such as network efficiency and node degree, across time windows of EEG recordings. The goal is to characterize differences between pathological and healthy groups at both the node and network levels. We particularly examine epilepsy and healthy subject groups to highlight differences in local network efficiency and node degrees, with statistical significance confirmed via F-tests.

We are concerned about the main encoding granule cells (GCs) in the hippocampal dentate gyrus (DG). Young immature GCs (imGCs) appear through adult neurogenesis. In comparison to the mature GCs (mGCs) (born during development), the imGCs show high activation due to lower firing threshold. On the other hand, they receive low excitatory drive from the entorhinal cortex via perforant paths and from the hilar mossy cells with lower connection probability pc (= 20 x %) (x : synaptic connectivity fraction; 0 [≤] x [≤] 1) than the mGCs with the connection probability pc (= 20 %). Thus, the effect of low excitatory innervation (reducing activation degree) for the imGCs counteracts the effect of their high excitability. We consider a spiking neural network for the DG, incorporating both the mGCs and the imGCs. With decreasing x from 1 to 0, we investigate the effect of young adult-born imGCs on the sparsely synchronized rhythms (SSRs) of the GCs (mGCs, imGC, and whole GCs). For each x, population and individual firing behaviors in the SSRs are characterized in terms of the amplitude measure [Formula] (X = m, im, w for the mGCs, the imGCs, and the whole GCs, respectively) (representing the population synchronization degree) and the random phase-locking degree [Formula] (characterizing the regularity of individual single-cell discharges), respectively. We also note that, for 0 [≤] x [≤] 1, the mGCs and the imGCs exhibit pattern separation (i.e., a process of transforming similar input patterns into less similar output patterns) and pattern integration (making association between patterns), respectively. Quantitative relationship between SSRs and pattern separation and integration is also discussed. PACS numbers87.19.lj, 87.19.lm, 87.19.lv

Integration of information across heterogeneous sources creates added scientific value. It is, however, a challenge to progress, often a barrier, to interoperate data, tools and models across spatial and temporal scales. Here we present a design template for coupling simulators operating at different scales and enabling co-simulation. We illustrate its functioning along a neuroscience example, in which individual regions of interest are simulated on the cellular level to address mechanistic questions, 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 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 validate them against multiscale experiments, thereby largely widening the explanatory power of computational models.

In this study, we model high-dimensional Electroencephalogram signals in sleep stages, using a general trainable network of Hopf oscillators. The proposed architecture has two components: a layer of oscillators with lateral connections, and a complex valued feedforward network with and without a hidden layer. The output of the Hopf oscillators, whose dynamics is described in the complex domain, is fed as input to the feedforward network and the output predicts the EEG signals. The proposed network is trained in two stages: in the 1st stage, intrinsic frequencies of the oscillators and the lateral connections are trained whereas, in the 2nd stage, the complex-valued feed-forward network is trained. Reconstruction error obtained when there is a hidden layer in the feedforward network is an order of magnitude smaller than that obtained without a hidden layer. Also, it has been shown that during testing our model is able to generate EEG signals, whose spectral properties accurately describe the corresponding target signals. In the first, the oscillators do not have any spatial localization, whereas in the second the oscillators are spatially distributed in a spherical shell geometry. The model described can be interpreted as a stepping stone towards a large-scale model of brain dynamics.

Epilepsy is a neurological disorder that affects over 50 million individuals worldwide. Today, the gold-standard treatment for those who are drug resistant, meaning that symptoms cannot be controlled with medication, is to surgically remove the seizure onset zone (SOZ), the area of the brain believed to cause seizures: the main symptom of epilepsy. Unfortunately, around 50% of drug resistant patients are not resective candidates, which can be attributed in part to poor SOZ localization. SOZ localization is a complex and lengthy procedure, requiring visual inspection and manual processing by human experts that first need to localize and isolate seizure events. The intracranial electroencephalography (iEEG) is a tool that records electrophysiological activity of the inner brain at different regions and depths, and provides critical information on the SOZ. However, iEEG data processing methodologies are not standardized, and practice and resources vary across hospitals and clinics. To assist human experts with systematic processing of iEEG data, we propose a data processing pipeline that generates graph representations of iEEG data. We evaluate 9 different graph representations of publicly available iEEG data from 25 patients with epilepsy with a graph neural network model trained to detect seizures. Our results suggest that graph representations of iEEG data that leverage electrode and functional connectivity features are powerful data structures to analyze and interpret iEEG data in the context of epilepsy. We anticipate that our data pipeline that provides a systematic processing of neural data with graphs can integrate other data modalities like neuroimaging data. Moreover, methods used in the data pipeline have potentials to apply to other neurological disorders such as Parkinsons disease or major depression disorder.

Information processing in the brain is thought to result from the coordination of large-scale neu-ronal activity, but this occurs on a fine neural circuit. Here, we investigate their relationship. First, we estimate monosynaptic connectivity by applying an advanced analysis method to spike trains recorded with high-density microelectrodes and confirm that the estimated neuronal wiring is largely consistent with neuroanatomical and neurophysiological evidence. Second, we simulate calcium imaging signals from the same dataset and confirm that the estimated functional connectivity is influenced by shared inputs and population synchronization on slower timescales. Notably, even with unrealistically fast calcium dynamics, the functional connectivity is only partially consistent with the monosynaptic connectivity. These findings suggest the complementary roles for monosynaptic and functional connectivity: the former provides circuit-level specificity, while the latter reflects emergent system-wide patterns of activity. We propose that an integrative approach combining both perspectives is essential for understanding circuit-level computation in the brain.

The human brain is arguably the most complex "machine" to ever exist. Its detailed functioning is yet to be fully understood, let alone modeled. Neurological processes have logical signal-processing aspects and biophysical aspects, and both affect the brain structure, functioning and adaptation. Mathematical approaches based on both information and graph theory have been extensively used in an attempt to approximate its biological functioning, along with Artificial Intelligence approaches inspired by its logical functioning. In this article, we present an approach to model some aspects of the brain learning and signal processing, mimicking the metastability and backpropagation found in the real brain while also accounting for neuroplasticity. Several simulations are carried out with this model, to demonstrate how dynamic neuroplasticity, neural inhibition and neurons migration can remodel the brain logical connectivity to syncronize signal processing and obtain target latencies. This work demonstrates the importance of dynamic logical and biophysical remodelling in brain plasticity.

Brain functional networks are essential for understanding functional connectome. Computing the temporal dependencies between the regions of brain activities of functional magnetic resonance imaging (fMRI) gives us the functional connectivity between the regions. The pairwise connectivities in matrix form correspond to the functional network (fNet), also referred to as a functional connectivity network (FCN). We start with analyzing a correlation matrix, which is an adjacency matrix of the FCN. In this work, we perform a case study of comparison of different analytical approaches in finding node-communities of the brain network. We use five different methods of community detection, out of which two methods are implemented on the network after filtering out the edges with weight below a predetermined threshold. We additionally compute and observe the following characteristics of the outcomes: (i) modularity of the communities, (ii) symmetrical node-partition between the left and right hemispheres of the brain, i.e., hemispheric symmetry, and (iii) hierarchical modular organization. Our contribution is in identifying an appropriate test-bed for comparison of outcomes of approaches using different semantics, such as network science, information theory, multivariate analysis, and data mining.

Large scale brain models encompassing cortico-cortical, thalamo-cortical and basal ganglia processing are fundamental to understand the brain as an integrated system in healthy and disease conditions but are complex to analyze and interpret. Neuronal processes are typically segmented by region and modality in order to explain an experimental observation at a given scale, but integrative frameworks linking scales and modalities are scarce. Here, we present a set of functional requirements used to evaluate the recently developed large-scale brain model against a learning task involving coordinated learning between cortical and sub-cortical systems. The original Information Based Exchange Brain model (IBEx) is decomposed into functionally relevant subsystems, and each subsystem is analyzed and tuned independently and with regard to its relevant functional requirements. Intermediate conclusions are made for each subsystems according to the constraints imposed by these requirements. Subsystems are then re-introduced into the global framework. The relationship between the global framework and phenotypes associated with Huntingtons disease is then discussed and the framework considered in the context of other state-of-the-art integrative brain models.

Metastability in the brain is thought to be a mechanism involved in dynamic organization of cognitive and behavioral functions across multiple spatiotemporal scales. However, it is not clear how such organization is realized in underlying neural oscillations in a high-dimensional state space. It was shown that macroscopic oscillations often form phase-phase coupling (PPC) and phase-amplitude coupling (PAC) which result in synchronization and amplitude modulation, respectively, even without external stimuli. These oscillations can also make spontaneous transitions across synchronous states at rest. Using resting-state electroencephalographic signals and the autism-spectrum quotient scores acquired from healthy humans, we show experimental evidence that the PAC combined with PPC allows amplitude modulation to be transient, and that the metastable dynamics with this transient modulation is associated with autistic-like traits. In individuals with a longer attention span, such dynamics tended to show fewer transitions between states by forming delta-alpha PAC. We identified these states as two-dimensional metastable states that could share consistent patterns across individuals. Our findings suggest that the human brain dynamically organizes inter-individual differences in a hierarchy of macroscopic oscillations with multiple timescales by utilizing metastability. Author SummaryThe human brain organizes cognitive and behavioral functions dynamically. For decades, the dynamic organization of underlying neural oscillations has been a fundamental topic in neuroscience research. Even without external stimuli, macroscopic oscillations often form phase-phase coupling and phase-amplitude coupling (PAC) that result in synchronization and amplitude modulation, respectively, and can make spontaneous transitions across synchronous states at rest. Using resting-state electroencephalography signals acquired from healthy humans, we show evidence that these two neural couplings enable amplitude modulation to be transient, and that this transient modulation can be viewed as the transition among oscillatory states with different PAC strengths. We also demonstrate that such transition dynamics are associated with the ability to maintain attention to detail and to switch attention, as measured by autism-spectrum quotient scores. These individual dynamics were visualized as a trajectory among states with attracting tendencies, and involved consistent brain states across individuals. Our findings have significant implications for unraveling variability in the individual brains showing typical and atypical development.

Studying connectivity of neuronal cultures can provide insights for understanding brain networks but it is challenging to reveal neuronal connectivity from measurements. We apply a novel method that uses a theoretical relation between the time-lagged cross-covariance and the equal-time cross-covariance to reveal directed effective connectivity and synaptic weights of cortical neuron cultures at different days in vitro from multielectrode array recordings. Using a stochastic leaky-integrate-and-fire model, we show that the simulated spiking activity of the reconstructed networks can well capture the measured network bursts. The neuronal networks are found to be highly nonrandom with an over-representation of bidirectionally connections as compared to a random network of the same connection probability, with the fraction of inhibitory nodes comparable to the measured fractions of inhibitory neurons in various cortical regions in monkey, and have small-world topology with basic network measures comparable to those of the nematode C. elegans chemical synaptic network. Our analyses further reveal that (i) the excitatory and inhibitory incoming degrees have bimodal distributions the excitatory and inhibitory incoming degrees have bimodal distributions, which are that distributions that have been indicated to be optimal against both random failures and attacks in undirected networks; (ii) the distribution of the physical length of excitatory incoming links has two peaks indicating that excitatory signal is transmitted at two spatial scales, one localized to nearest nodes and the other spatially extended to nodes millimeters away, and the shortest links are mostly excitatory towards excitatory nodes and have larger synaptic weights on average; (iii) the average incoming and outgoing synaptic strength is non-Gaussian with long tails and, in particular, the distribution of outgoing synaptic strength of excitatory nodes with excitatory incoming synaptic strength is lognormal, similar to the measured excitatory postsynaptic potential in rat cortex. Author summaryTo understand how the brain processes signal and carries out its function, it is useful to know the connectivity of the underlying neuronal circuits. For large-scale neuronal networks, it is difficult to measure connectivity directly using electron microscopy techniques and methods that can estimate connectivity from electrophysiological recordings are thus highly desirable. Existing methods focus mainly on estimating functional connectivity, which is defined by statistical dependencies between neuronal activities but the relevant direct casual interactions are captured by effective connectivity. Here we apply a novel covariance-relation based method to estimate the directed effective connectivity and synaptic weights of cortical neuron cultures from recordings of multielectrode array of over 4000 electrodes taken at different days in vitro. The neuronal networks are found to be nonrandom, small-world, excitation/inhibition balanced as measured in monkey cortex, and with feeder hubs. Our analyses further suggest some form of specialisation of nodes in receiving excitatory and inhibitory signals and the transmission of excitatory signals at two spatial scales, one localized to nearest nodes and the other spatially extended to nodes millimeters away, and reveal that the distributions of the average incoming and outgoing synaptic strength are skewed with long tails.

Epilepsy is a common neurological disorder that sub-stantially deteriorates patients safety and quality of life. Electroencephalogram (EEG) has been the golden-standard technique for diagnosing this brain disorder and has played an essential role in epilepsy monitoring and disease management. It is extremely laborious and challenging, if not practical, for physicians and expert humans to annotate all recorded signals, particularly in long-term monitoring. The annotation process often involves identifying signal segments with suspected epileptic seizure features or other abnormalities and/or known healthy features. Therefore, automated epilepsy detection becomes a key clinical need because it can greatly improve clinical practices efficiency and free up human expert time to attend to other important tasks. Current automated seizure detection algorithms generally face two challenges: (1) models trained for specific patients, but such models are patient-specific, hence fail to generalize to other patients and real-world situations; (2) seizure detection models trained on large EEG datasets have low sensitivity and/or high false positive rates, often with an area under the receiver operating characteristic (AUROC) that is not high enough for potential clinical applicability. This paper proposes Transformers for Seizure Detection, which we refer to as TSD in this manuscript. A Transformer is a deep learning architecture based on an encoder-decoder structure and on attention mechanisms, which we apply to recorded brain signals. The AUROC of our proposed model has achieved 92.1%, tested with Temple Universitys publically available electroencephalogram (EEG) seizure corpus dataset (TUH). Additionally, we highlight the impact of input domains on the models performance. Specifically, TSD performs best in identifying epileptic seizures when the input domain is a time-frequency. Finally, our proposed model for seizure detection in inference-only mode with EEG recordings shows outstanding performance in classifying seizure types and superior model initialization.

The Virtual Epileptic Patient (VEP) refers to a computer-based representation of a patient with epilepsy that combines personalized anatomical data with dynamical models of abnormal brain activities. It is capable of generating spatio-temporal seizure patterns that resemble those recorded with invasive methods such as stereoelectro EEG data, allowing for the evaluation of clinical hypotheses before planning surgery. This study highlights the effectiveness of calibrating VEP models using a global optimization approach. The approach utilizes SaCeSS, a cooperative metaheuristic capable of parallel computation, to yield high-quality solutions without requiring excessive computational time. Through extensive benchmarking, our proposal successfully solved a set of different configurations of VEP models, demonstrating better scalability and superior performance against other parallel solvers. These results were further enhanced using a Bayesian optimization framework for hyperparameter tuning, with significant gains in terms of both accuracy and computational cost. Additionally, we added an scalable uncertainty quantification phase after model calibration, and used it to assess the variability in estimated parameters across different problems. Overall, this study has the potential to improve the estimation of pathological brain areas in drug-resistant epilepsy, thereby to inform the clinical decision-making process. Author summaryMotivated by the problem of parameter estimation in a set of whole-brain network models of epilepsy (of increasing complexity), this study addresses the question of choosing a robust global optimization solver that can be accelerated by exploiting parallelism in different infrastructures, from desktop workstations to supercomputers. By leveraging data-driven techniques with robust cooperative global optimization methods, we aim to achieve accurate parameter estimation with reduced reliance on prior information. This is due to the dependency of Bayesian inference on the level of information in the prior, while this approach allows us to quantify uncertainty in the absence of any prior knowledge effectively. In this work, we construct an efficient and accurate method to perform parameter estimation and uncertainty quantification for the VEP model, and we use it to infer the brain regional epileptogenicity from source and sensor level whole-brain data. Of specific interest is the ability of our method to produce inference for high-dimensional state-space models governed by deterministic, stochastic, well-behaved, and stiff differential equations, using only partial observations and sparse encoding from system states to the observation.

The development of biologically realistic models of brain microcircuits and regions is currently a very relevant topic in computational neuroscience. From basic research to clinical applications, there is an increasing demand for accurate models that incorporate local cellular and network specificities, able to capture a broad range of dynamics and functions associated with given brain regions. One of the main challenges of these models is the passage between different scales, going from the microscale (cellular) to the meso (microcircuit) and macroscale (region or whole-brain level), while keeping at the same time a constraint on the demand of computational resources. One novel approach to this problem is the use of mean-field models of neuronal activity to build large-scale simulations. This provides an effective solution to the passage between scales with relatively low computational demands, which is achieved by a drastic reduction in the dimensionality of the system. In this paper we introduce a multiscale modelling framework for the hippocampal CA1, a region of the brain that plays a key role in functions such as learning, memory consolidation and navigation. Our modelling framework goes from the single cell level to the macroscale and makes use of a novel mean-field model of CA1, introduced in this paper, to bridge the gap between the micro and macro scales. To develop the mean-field model we make use of a recently introduced formalism based on a bottom-up approach that is easily applicable to different neuronal models and cell types. We test and validate the model by analyzing the response of the system to the main brain rhythms observed in the hippocampus and comparing our results with the ones of the corresponding spiking network model of CA1. In addition, we show an example of the implementation of our model to study a stimulus propagation at the macro-scale, and we compare the results obtained from our model with the corresponding spiking network model of the whole CA1 area.

Conventional deep neural networks capture essential information processing stages in perception. Deep neural networks often require very large volume of training examples, whereas children can learn concepts such as hand-written digits with few examples. The goal of this project is to develop a deep spiking neural network that can learn from few training trials. Using known neuronal mechanisms, a spiking neural network model is developed and trained to recognize hand-written digits with presenting one to four training examples for each digit taken from the MNIST database. The model detects and learns geometric features of the images from MNIST database. In this work, a novel biological back-propagation based learning rule is developed and used to a train the network to detect basic features of different digits. For this purpose, randomly initialized synaptic weights between the layers are being updated. By using a neuroscience inspired mechanism named synaptic pruning and a predefined threshold, some of the synapses through the training are deleted. Hence, information channels are constructed that are highly specific for each digit as matrix of synaptic connections between two layers of spiking neural networks. These connection matrixes named information channels are used in the test phase to assign a digit class to each test image. As similar to humans abilities to learn from small training trials, the developed spiking neural network needs a very small dataset for training, compared to conventional deep learning methods checked on MNIST dataset.