Large-scale Network Models Research Articles

Adapting magnetoresistive memory devices for accurate and on-chip-training-free in-memory computing.

Memristors have emerged as promising devices for enabling efficient multiply-accumulate (MAC) operations in crossbar arrays, crucial for analog in-memory computing (AiMC). However, variations in memristors and associated circuits can affect the accuracy of analog computing. Typically, this is mitigated by on-chip training, which is challenging for memristors with limited endurance. We present a hardware-software codesign using magnetic tunnel junction (MTJ)-based AiMC off-chip calibration that achieves software accuracy without costly on-chip training. Hardware-wise, MTJ devices exhibit ultralow cycle-to-cycle variations, as experimentally evaluated over 1 million mass-produced devices. Software-wise, leveraging this, we propose an off-chip training method to adjust deep neural network parameters, achieving accurate AiMC inference. We validate this approach with MAC operations, showing improved transfer curve linearity and reduced errors. By emulating large-scale neural network models, our codesigned MTJ-based AiMC closely matches software baseline accuracy and outperforms existing off-chip training methods, highlighting MTJ's potential in AI tasks.

Utilising activity patterns of a complex biophysical network model to optimise intra-striatal deep brain stimulation

A large-scale biophysical network model for the isolated striatal body is developed to optimise potential intrastriatal deep brain stimulation applied to, e.g. obsessive-compulsive disorder. The model is based on modified Hodgkin–Huxley equations with small-world connectivity, while the spatial information about the positions of the neurons is taken from a detailed human atlas. The model produces neuronal spatiotemporal activity patterns segregating healthy from pathological conditions. Three biomarkers were used for the optimisation of stimulation protocols regarding stimulation frequency, amplitude and localisation: the mean activity of the entire network, the frequency spectrum of the entire network (rhythmicity) and a combination of the above two. By minimising the deviation of the aforementioned biomarkers from the normal state, we compute the optimal deep brain stimulation parameters, regarding position, amplitude and frequency. Our results suggest that in the DBS optimisation process, there is a clear trade-off between frequency synchronisation and overall network activity, which has also been observed during in vivo studies.

Abstract We135: Mechanistic Modeling Identifies Novel Regulators of Cardiac Hypertrophy

Introduction: Cardiac hypertrophy is a leading predictor of heart failure. The mechanisms that drive cardiac hypertrophy are poorly understood from a molecular and cellular perspective, due to the lack of knowledge of the governing signaling. Hypothesis and Aims: I hypothesize that there exist genes that regulate hypertrophy beyond the field’s current understanding that can be identified by combining the publicly available, high-throughput International Mouse Phenotyping Consortium (IMPC) database with network modeling. I aim to use a large-scale network model of cardiac hypertrophy to assess the IMPC genes’ effects on hypertrophic phenotypes and validate these findings in vitro to identify novel regulators. Materials & Methods: The IMPC was established by worldwide mouse clinics using in vivo models to perform single gene knockout experiments with primary phenotyping, forming a database with ~9000 genes, ~1000 of which can be associated with cardiac hypertrophy. 1 st neighbors (genes directly interacting with the hypertrophy network model as evidenced by protein-protein interaction databases) were identified and effects on hypertrophy were assessed through the IMPC and network model knockdown simulations. When both methods suggested hypertrophic regulation, a gene was considered a candidate regulator. siRNA experiments are being performed in neonatal rat cardiomyocytes for validation. High-content imaging and qPCR are used to quantify changes in phenotype and confirm gene knockdown. Results: Combining the IMPC with protein-protein interaction data, 37 1 st neighbor genes were identified. Knockdown simulation identified 2 genes (LRIG1, CBL) as negative regulators. 3 genes (VAV2, LEPR, RASA1) were identified as positive regulators. These genes became candidates for in vitro validation given the network-wide hypertrophic effects enacted by their knockdowns. siRNA experiments are being optimized to fully validate modeling findings. Conclusions: Mechanistic modeling identified 5 genes as novel regulators of cardiac hypertrophy. Sensitivity analysis highlighted known influencers of hypertrophy in the knockdown context. siRNA knockdown experiments are being performed to validate the model predictions.

Emergent effects of synaptic connectivity on the dynamics of global and local slow waves in a large-scale thalamocortical network model of the human brain.

Slow-wave sleep (SWS), characterized by slow oscillations (SOs, <1Hz) of alternating active and silent states in the thalamocortical network, is a primary brain state during Non-Rapid Eye Movement (NREM) sleep. In the last two decades, the traditional view of SWS as a global and uniform whole-brain state has been challenged by a growing body of evidence indicating that SO can be local and can coexist with wake-like activity. However, the mechanisms by which global and local SOs arise from micro-scale neuronal dynamics and network connectivity remain poorly understood. We developed a multi-scale, biophysically realistic human whole-brain thalamocortical network model capable of transitioning between the awake state and SWS, and we investigated the role of connectivity in the spatio-temporal dynamics of sleep SO. We found that the overall strength and a relative balance between long and short-range synaptic connections determined the network state. Importantly, for a range of synaptic strengths, the model demonstrated complex mixed SO states, where periods of synchronized global slow-wave activity were intermittent with the periods of asynchronous local slow-waves. An increase in the overall synaptic strength led to synchronized global SO, while a decrease in synaptic connectivity produced only local slow-waves that would not propagate beyond local areas. These results were compared to human data to validate probable models of biophysically realistic SO. The model producing mixed states provided the best match to the spatial coherence profile and the functional connectivity estimated from human subjects. These findings shed light on how the spatio-temporal properties of SO emerge from local and global cortical connectivity and provide a framework for further exploring the mechanisms and functions of SWS in health and disease.

Complex-valued sensitivity analysis tool aimed to power flow optimization

Abstract In this paper, the use of complex differentiation is presented as a robust and concise tool for modeling and sensitivity analysis of large-scale electrical networks. A complex power flow model was employed to determine the network’s operating point, from which the sensitivity analysis of the electrical system is conducted. The proposed methodology, utilizing complex differentiation, was applied to two systems with data extracted from the MATPOWER software. The obtained results are depicted in a network topology graph, highlighting the identified sensitivities.

Iterative algorithm computational spectrometer based on a single-hidden-layer neural network

Computational spectrometers have great application prospects in hyperspectral detection, and fast and high-precision in situ measurement is an important development trend. The computational spectrometer based on iterative algorithms has low requirements for computational resources and is easy to achieve hardware integration and in situ measurement. However, iterative algorithms are difficult to achieve high reconstruction accuracy due to the ill-posed nature of problems. Neural networks have powerful learning capabilities and can achieve high-precision spectral reconstruction. However, solely relying on neural network algorithms for reconstruction requires higher storage space and computing power from hardware devices, which makes it difficult to integrate large-scale neural network models into embedded systems. We propose using neural networks to alleviate the effect of the problem ill-posedness on the reconstruction results of iterative algorithms, so as to improve the reconstruction accuracy of the iterative algorithm computational spectrometers. First, spectral reconstruction was performed with iterative algorithms using a public spectral dataset. Then, a single-hidden-layer neural network was trained to establish a fitting relationship between the iterative algorithm spectral reconstruction results and the original spectrum. Finally, simulation and experimental results show that the proposed application of neural networks to alleviate the ill-posed problem of the iterative algorithm spectral reconstruction can effectively improve the reconstruction accuracy of iterative algorithm computational spectrometers with low computational resources. The research results may have good potential in achieving fast and high-precision in situ measurements of computational spectrometers.

A large-scale neuronal network modelling study: Stimulus size modulates gamma oscillations in the primary visual cortex by long-range connections.

Stimulus size modulation of neuronal firing activity is a fundamental property of the primary visual cortex. Numerous biological experiments have shown that stimulus size modulation is affected by multiple factors at different spatiotemporal scales, but the exact pathways and mechanisms remain incompletely understood. In this paper, we establish a large-scale neuronal network model of primary visual cortex withlayer 2/3 to study how gamma oscillation properties are modulated by stimulus size and especially how long-range connections affect the modulation as realistic neuronal properties and spatial distributions of synaptic connections are considered. It is shown that long-range horizontal synaptic connections are sufficient to produce dimensional modulation of firing rates and gamma oscillations. In particular, with increasing grating stimulus size, the firing rate increases and then decreases, the peak frequency of gamma oscillations decreases and the spectral power increases. These are consistent with biological experimental observations. Furthermore, we explain in detail how the number and spatial distribution of long-range connections affect the size modulation of gamma oscillations by using the analysis of neuronal firing activity and synaptic current fluctuations. Our results provide a mechanism explanation for size modulation of gamma oscillations in the primary visual cortex and reveal the important and unique role played by long-range connections, which contributes to a deeper understanding of the cognitive function of gamma oscillations in visual cortex.

Emergent effects of synaptic connectivity on the dynamics of global and local slow waves in a large-scale thalamocortical network model of the human brain.

Slow-wave sleep (SWS), characterized by slow oscillations (SO, <1Hz) of alternating active and silent states in the thalamocortical network, is a primary brain state during Non-Rapid Eye Movement (NREM) sleep. In the last two decades, the traditional view of SWS as a global and uniform whole-brain state has been challenged by a growing body of evidence indicating that SO can be local and can coexist with wake-like activity. However, the understanding of how global and local SO emerges from micro-scale neuron dynamics and network connectivity remains unclear. We developed a multi-scale, biophysically realistic human whole-brain thalamocortical network model capable of transitioning between the awake state and slow-wave sleep, and we investigated the role of connectivity in the spatio-temporal dynamics of sleep SO. We found that the overall strength and a relative balance between long and short-range synaptic connections determined the network state. Importantly, for a range of synaptic strengths, the model demonstrated complex mixed SO states, where periods of synchronized global slow-wave activity were intermittent with the periods of asynchronous local slow-waves. Increase of the overall synaptic strength led to synchronized global SO, while decrease of synaptic connectivity produced only local slow-waves that would not propagate beyond local area. These results were compared to human data to validate probable models of biophysically realistic SO. The model producing mixed states provided the best match to the spatial coherence profile and the functional connectivity estimated from human subjects. These findings shed light on how the spatio-temporal properties of SO emerge from local and global cortical connectivity and provide a framework for further exploring the mechanisms and functions of SWS in health and disease.

GAT TransPruning: progressive channel pruning strategy combining graph attention network and transformer.

Recently, large-scale artificial intelligence models with billions of parameters have achieved good results in experiments, but their practical deployment on edge computing platforms is often subject to many constraints because of their resource requirements. These models require powerful computing platforms with a high memory capacity to store and process the numerous parameters and activations, which makes it challenging to deploy these large-scale models directly. Therefore, model compression techniques are crucial role in making these models more practical and accessible. In this article, a progressive channel pruning strategy combining graph attention network and transformer, namely GAT TransPruning, is proposed, which uses the graph attention networks (GAT) and the attention of transformer mechanism to determine the channel-to-channel relationship in large networks. This approach ensures that the network maintains its critical functional connections and optimizes the trade-off between model size and performance. In this study, VGG-16, VGG-19, ResNet-18, ResNet-34, and ResNet-50 are used as large-scale network models with the CIFAR-10 and CIFAR-100 datasets for verification and quantitative analysis of the proposed progressive channel pruning strategy. The experimental results reveal that the accuracy rate only drops by 6.58% when the channel pruning rate is 89% for VGG-19/CIFAR-100. In addition, the lightweight model inference speed is 9.10 times faster than that of the original large model. In comparison with the traditional channel pruning schemes, the proposed progressive channel pruning strategy based on the GAT and Transformer cannot only cut out the insignificant weight channels and effectively reduce the model size, but also ensure that the performance drop rate of its lightweight model is still the smallest even under high pruning ratio.

Automatic processing of dynamic responses via wavelet transform for the development of dynamic equivalent models

The ever-increasing requirements for electricity, the emergence of microgrids and the escalating penetration of distributed generators has reinvigorated the interest in dynamic equivalencing of distribution networks, due to its significance in power system analysis. Scope of this paper is to propose an efficient method for event detection and filtering of dynamic responses based on the wavelet transform (WT), in order to improve the quality of signals used for the derivation of dynamic equivalent models. The accuracy of the proposed method is tested using artificially created noisy responses, by applying Gaussian, Laplace, and Student’s-t noise distributions. Comparisons with other filtering techniques are also performed and the impact of all methods on the derivation of accurate equivalent model parameters is quantified and analyzed. The performance of the proposed method is also evaluated by using RMS responses obtained from a large-scale distribution network model as well as by analyzing laboratory measurements; results verify the efficiency and applicability of the WT-based processing procedure, by achieving a parameter estimation error well below 1%. It is noteworthy that the average computational burden throughout the process remains under 0.39 s.

Mathematical generation of data-driven hippocampal CA1 pyramidal neurons and interneurons copies via A-GLIF models for large-scale networks covering the experimental variability range

Efficient and accurate large-scale networks are a fundamental tool in modeling brain areas, to advance our understanding of neuronal dynamics. However, their implementation faces two key issues: computational efficiency and heterogeneity. Computational efficiency is achieved using simplified neurons, whereas there are no practical solutions available to solve the problem of reproducing in a large-scale network the experimentally observed heterogeneity of the intrinsic properties of neurons. This is important, because the use of identical nodes in a network can generate artifacts which can hinder an adequate representation of the properties of a real network.To this aim, we introduce a mathematical procedure to generate an arbitrary large number of copies of simplified hippocampal CA1 pyramidal neurons and interneurons models, which exhibit the full range of firing dynamics observed in these cells — including adapting, non-adapting and bursting. For this purpose, we rely on a recently published adaptive generalized leaky integrate-and-fire (A-GLIF) modeling approach, leveraging on its ability to reproduce the rich set of electrophysiological behaviors of these types of neurons under a variety of different stimulation currents.The generation procedure is based on a perturbation of model’s parameters related to the initial data, firing block, and internal dynamics, and suitably validated against experimental data to ensure that the firing dynamics of any given cell copy remains within the experimental range. A classification procedure confirmed that the firing behavior of most of the pyramidal/interneuron copies was consistent with the experimental data. This approach allows to obtain heterogeneous copies with mathematically controlled firing properties. A full set of heterogeneous neurons composing the CA1 region of a rat hippocampus (approximately 1.2 million neurons), are provided in a database freely available in the live paper section of the EBRAINS platform.By adapting the underlying A-GLIF framework, it will be possible to extend the numerical approach presented here to create, in a mathematically controlled manner, an arbitrarily large number of non-identical copies of cell populations with firing properties related to other brain areas.

Cortical cell assemblies and their underlying connectivity: An in silico study.

Recent developments in experimental techniques have enabled simultaneous recordings from thousands of neurons, enabling the study of functional cell assemblies. However, determining the patterns of synaptic connectivity giving rise to these assemblies remains challenging. To address this, we developed a complementary, simulation-based approach, using a detailed, large-scale cortical network model. Using a combination of established methods we detected functional cell assemblies from the stimulus-evoked spiking activity of 186,665 neurons. We studied how the structure of synaptic connectivity underlies assembly composition, quantifying the effects of thalamic innervation, recurrent connectivity, and the spatial arrangement of synapses on dendrites. We determined that these features reduce up to 30%, 22%, and 10% of the uncertainty of a neuron belonging to an assembly. The detected assemblies were activated in a stimulus-specific sequence and were grouped based on their position in the sequence. We found that the different groups were affected to different degrees by the structural features we considered. Additionally, connectivity was more predictive of assembly membership if its direction aligned with the temporal order of assembly activation, if it originated from strongly interconnected populations, and if synapses clustered on dendritic branches. In summary, reversing Hebb's postulate, we showed how cells that are wired together, fire together, quantifying how connectivity patterns interact to shape the emergence of assemblies. This includes a qualitative aspect of connectivity: not just the amount, but also the local structure matters; from the subcellular level in the form of dendritic clustering to the presence of specific network motifs.

Experimental and numerical analysis of flow through a natural rough fracture subject to normal loading

Fractured crystalline rocks have been chosen or are under consideration by several countries as host rock formations for deep geological repositories for spent nuclear fuel. In such geological formations, flow and solute transport are mostly controlled by a network of connected natural fractures, each of them being characterised by internal heterogeneity, also denoted as roughness. Fractures are, in turn, subject to variable load caused by various factors, such as the presence of thick ice sheets formed during glaciation periods. Understanding how coupled hydro-mechanical (HM) processes affect flow and transport at the scale of a single natural fracture is crucial for a robust parameterisation of large-scale discrete fracture network models, which are not only used for nuclear waste disposal applications but are also of interest to problems related to geothermics, oil and gas production or groundwater remediation. In this work, we analyse and model an HM experiment carried out in a single natural fracture and use the results of both, the experimental and the modelling work, to get insights into fundamental questions such as the applicability of local cubic law or the effect of normal load on channeling. The initial fracture aperture was obtained from laser scanning of the two fracture surfaces and an equivalent initial aperture was then defined by moving the two fracture surfaces together and comparing the results obtained using a Navier–Stokes based computational fluid dynamics (CFD) model with the experimental flowrate obtained for unloaded conditions. The mechanical effect of the different loading stages was simulated using a high-resolution contact model. The different computed fracture apertures were then used to run groundwater flow simulations using a modified Reynolds equation. The results show that, without correction, local cubic law largely overestimates flowrates. Instead, we show that by explicitly acknowledging the difference between the mechanical aperture and the hydraulic aperture and setting the latter equal to 1/5 of the former, cubic law provides a very reasonable approximation of the experimental flowrates over the entire loading cycle. A positive correlation between fluid flow channeling and normal load is also found.

Large-Scale Bio-Inspired FPGA Models for Path Planning.

The hippocampus provides significant inspiration for spatial navigation and memory in both humans and animals. Constructing large-scale spiking neural network (SNN) models based on the biological neural systems is an important approach to comprehend the computational principles and cognitive function of the hippocampus. Such models are usually implemented on neuromorphic computing platforms, which often have limited computing resources that constrain the achievable scale of the network. This work introduces a series of digital design methods to realize a Field-Programmable Gate Array (FPGA) friendly SNN model. The methods include FPGA-friendly nonlinear calculation modules and a fixed-point design algorithm. A brain-inspired large-scale SNN of ∼21 k place cells for path planning is mapped on FPGA. The results show that the path planning tasks in different environments are finished in real-time and the firing activities of place cells are successfully reproduced. With these methods, the achievable network size on one FPGA chip is increased by 1595 times with higher resource usage efficiency and faster computation speed compared to the state-of-the-art.

NoC simulation steered by NEST: McAERsim and a Noxim patch.

Great knowledge was gained about the computational substrate of the brain, but the way in which components and entities interact to perform information processing still remains a secret. Complex and large-scale network models have been developed to unveil processes at the ensemble level taking place over a large range of timescales. They challenge any kind of simulation platform, so that efficient implementations need to be developed that gain from focusing on a set of relevant models. With increasing network sizes imposed by these models, low latency inter-node communication becomes a critical aspect. This situation is even accentuated, if slow processes like learning should be covered, that require faster than real-time simulation. Therefore, this article presents two simulation frameworks, in which network-on-chip simulators are interfaced with the neuroscientific development environment NEST. This combination yields network traffic that is directly defined by the relevant neural network models and used to steer the network-on-chip simulations. As one of the outcomes, instructive statistics on network latencies are obtained. Since time stamps of different granularity are used by the simulators, a conversion is required that can be exploited to emulate an intended acceleration factor. By application of the frameworks to scaled versions of the cortical microcircuit model-selected because of its unique properties as well as challenging demands-performance curves, latency, and traffic distributions could be determined. The distinct characteristic of the second framework is its tree-based source-address driven multicast support, which, in connection with the torus topology, always led to the best results. Although currently biased by some inherent assumptions of the network-on-chip simulators, the results suit well to those of previous work dealing with node internals and suggesting accelerated simulations to be in reach.

Visual Diagnostics of Parallel Performance in Training Large-Scale DNN Models.

Diagnosing the cluster-based performance of large-scale deep neural network (DNN) models during training is essential for improving training efficiency and reducing resource consumption. However, it remains challenging due to the incomprehensibility of the parallelization strategy and the sheer volume of complex data generated in the training processes. Prior works visually analyze performance profiles and timeline traces to identify anomalies from the perspective of individual devices in the cluster, which is not amenable for studying the root cause of anomalies. In this article, we present a visual analytics approach that empowers analysts to visually explore the parallel training process of a DNN model and interactively diagnose the root cause of a performance issue. A set of design requirements is gathered through discussions with domain experts. We propose an enhanced execution flow of model operators for illustrating parallelization strategies within the computational graph layout. We design and implement an enhanced Marey's graph representation, which introduces the concept of time-span and a banded visual metaphor to convey training dynamics and help experts identify inefficient training processes. We also propose a visual aggregation technique to improve visualization efficiency. We evaluate our approach using case studies, a user study and expert interviews on two large-scale models run in a cluster, namely, the PanGu- α 13B model (40 layers), and the Resnet model (50 layers).

An interpretable mechanism for grating-induced cross-inhibition and gamma oscillation based on a visual cortical neuronal network model

&lt;abstract&gt;&lt;p&gt;Biological experiments targeting the mammalian primary visual cortex have shown that neuronal response to a preferred orientation grating is cross-inhibited by an orthogonal orientation mask grating. The plaid formed by the overlap of the two gratings not only causes a decrease in the neuronal firing rate but also shifts the gamma oscillation to a weaker oscillation at a higher frequency. The mechanism for the above phenomena is unclarified. In this paper, a large-scale cortical neuronal network model with biological details is constructed. In this model, two modes of connectivity that may contribute to cross-inhibition are considered: the thalamo-cortical feedforward pathway and the push-pull organization of cortical layer 4. Based on this model, the modulation of firing rate and gamma oscillation by a plaid stimulation are successfully reproduced, which is consistent with biological experiments and suggests that it is the thalamo-cortical feedforward pathway that leads to cross-inhibition. Furthermore, our analysis of the neuronal spike clusters and current fluctuations suggests that the push-pull organization leads to an increase in gamma frequency during the transition of visual stimuli from grating to plaid by modulating the source of synaptic inhibition in local neuronal populations. Such results will help to understand the visual processing under multi-input integration.&lt;/p&gt;&lt;/abstract&gt;

Stability and Dynamics Analysis of Time-Delay Fractional-Order Large-Scale Dual-Loop Neural Network Model With Cross-Coupling Structure.

In recent years, the analysis of the dynamics of annular neural networks has received extensive attention and achieved some achievements. However, most of the current research merely focuses on the single-ring, low-dimension, two rings sharing one neuron cases, without considering the rich coupling modes between rings. In this article, a large-scale time-delay fractional-order dual-loop neural network model with cross-coupling structure is established, in which two rings complete information interaction through two shared neurons. Moreover, the Caputo fractional derivative is introduced in this article to describe the neural network more accurately. First, the transmission time delay between each neuron is selected as the key parameter leading to the bifurcation, and the characteristic equation of the network is creatively derived using the Coates flow graph method. Subsequently, through the holistic element method and magnitude angle formula, we simplify the analytical process. Then, we obtain the stability and Hopf bifurcation criterion of the network. Finally, the conclusions of the theoretical analysis are verified by a series of numerical simulations. The results show that the stability region of the network is closely related to the fractional order, the number of neurons, the distribution of neurons, and the self-feedback coefficients. Moreover, the time delays have a significant effect on the amplitude and period of the Hopf bifurcation.

ADGNN: Towards Scalable GNN Training with Aggregation-Difference Aware Sampling

Distributed computing is promising to enable large-scale graph neural network (GNN) model training. However, care is needed to avoid excessive computational and communication overheads. Sampling is promising in terms of enabling scalability, and sampling techniques have been proposed to reduce training costs. However, online sampling introduces large overheads, and while offline sampling that is done only once can eliminate such overheads, it instead introduces information loss and accuracy degradation. Thus, existing sampling techniques are unable to improve simultaneously both efficiency and accuracy, particularly at low sampling rates. We develop a distributed system, ADGNN, for full-batch based GNN training that adopts a hybrid sampling architecture to enable a trade-off between efficiency and accuracy. Specifically, ADGNN employs sampling result reuse techniques to reduce the cost associated with sampling and thus improve training efficiency. To alleviate accuracy degradation, we introduce a new metric,Aggregation Difference (AD), that quantifies the gap between sampled and full neighbor set aggregation. We present so-called AD-Sampling that aims to minimize the Aggregation Difference with an adaptive sampling frequency tuner. Finally, ADGNN employs anAD -importance-based sampling technique for remote neighbors to further reduce communication costs. Experiments on five real datasets show that ADGNN is able to outperform the state-of-the-art by up to nearly 9 times in terms of efficiency, while achieving comparable accuracy to the non-sampling methods.

Invited Session I: 30 years of normalization in the visual system: Spatiotemporal fluctuations in E/I balance & their impact on perception.

The Normalization Model has helped illuminate our understanding of neural computations in a wide variety of contexts. Here, I will describe our recent efforts to understand how intrinsic fluctuations in the ratio of excitatory to inhibitory conductances play out in space and time, and how these fluctuations impact perception. We find that spontaneous cortical activity is organized into traveling waves that traverse visual cortex several times per second. Recording in Area MT of the common marmoset, we find that these intrinsic traveling waves (iTWs) regulate both the gain of the stimulus-evoked spiking response and the monkey's perceptual sensitivity. A large-scale spiking network model recapitulates the properties of iTWs measured in vivo and explains the observed fluctuations in response gain as resulting from momentary fluctuations in E/I balance. The model predicts that iTWs are sparse, in the sense that only a small fraction of the neural population participates in any individual iTW. As a result, iTWs can occur without inducing correlated variability, which we have found, in separate experiments, can impair sensory discrimination. We thus refer to the model as the sparse-wave model of iTWs. Taken together, these findings lead to the conclusion that traveling waves strongly modulate neural and perceptual sensitivity.