Individual Spike Trains Research Articles

Neuroplastic alterations in common synaptic inputs and synergistic motor unit clusters controlling the vastii muscles of individuals with ACL reconstruction.

This cross-sectional study aims to elucidate the neural mechanisms underlying the control of knee extension forces in individuals with anterior cruciate ligament reconstruction (ACLR). Eleven soccer players with ACLR and nine control players performed unilateral isometric knee extensions at 10% and 30% of their maximum voluntary force (MVF). Simultaneous recordings of high-density surface electromyography (HDEMG) and force output were conducted for each lower limb, and HDEMG data from the vastus lateralis (VL) and vastus medialis (VM) muscles were decomposed into individual motor unit spike trains. Force steadiness was estimated using the coefficient of variation of force. An intramuscular coherence analysis was adopted to estimate the common synaptic input (CSI) converging to each muscle. A factor analysis was applied to investigate the neural strategies underlying the control of synergistic motor neuron clusters, referred to as motor unit modes. Force steadiness was similar between lower limbs. However, motor neurons innervating the VL on the reconstructed side received a lower proportion of CSI at low-frequency bandwidths (<5 Hz) compared with the unaffected lower limbs (P < 0.01). Furthermore, the reconstructed side demonstrated a higher proportion of motor units associated with the neural input common to the synergistic muscle, as compared with the unaffected lower limbs (P < 0.01). These findings indicate that the VL muscle of reconstructed lower limbs contribute marginally to force steadiness and that a plastic rearrangement in synergistic clusters of motor units involved in the control of knee extension forces is evident following ACLR.NEW & NOTEWORTHY Chronic quadriceps dysfunction is common after anterior cruciate ligament reconstruction (ACLR). We investigated voluntary force control strategies by estimating common inputs to motor neurons innervating the vastii muscles. Our results showed attenuated common inputs to the vastus lateralis and plastic rearrangements in functional clusters of motor neurons modulating knee extension forces in the reconstructed limb. These findings suggest neuroplastic adjustments following ACLR that may occur to fine-tune the control of quadriceps forces.

Common synaptic input between motor units from the lateral and medial posterior soleus compartments does not differ from that within each compartment.

The human soleus muscle is anatomically divided into four separate anatomical compartments. The functional role of this compartmentalization remains unclear. Here, we tested the hypothesis that the common synaptic input to motor units between the medial and lateral posterior compartments is less than within each compartment. Fourteen male participants performed three different heel-raise tasks that were considered to place a different mechanical demand on the medial and lateral soleus compartments. High-density electromyography (EMG) signals from the medial and lateral soleus compartments and the medial gastrocnemius of the right leg were decomposed into individual motor unit spike trains. The coherence between cumulative spike trains of the motor units was estimated. The coherence analysis was also repeated for motor units that were matched across all three tasks. Furthermore, we calculated the ratio of significant correlations between the spike trains of pairs of motor units. We observed that the coherence between motor units of the two soleus compartments was similar as the coherence between motor units within each compartment, regardless of the task. The correlation analysis performed on pairs of motor units confirmed these results. We conclude that the level of common synaptic input between the motor units innervating the medial and lateral posterior soleus compartment is not different than the common synaptic input between motor units innervating each of these compartments, which contrasts with findings from previous studies on finger muscles. This suggests that there is no independent neural control for the individual posterior soleus compartments.NEW & NOTEWORTHY The human soleus muscle is anatomically subdivided into four compartments. The functional role for this compartmentalization remains unknown. Here, we showed that, contrary to previous findings in finger muscles, the common synaptic input between motor units innervating the medial and lateral posterior soleus compartment was similar as that between motor units within the individual compartments. We suggest that the contradictory findings with other compartmentalized muscles may be explained by differences in muscle-tendon anatomy and function.

Acute Intermittent Hypoxia Increases Motor Unit Discharge Rates And Strength in Persons with SCI

<h3>Research Objectives</h3> To determine whether changes in motor unit discharge rates contribute to acute intermittent hypoxia (AIH)-induced improvements in volitional strength in persons with chronic incomplete spinal cord injury (iSCI) at the cervical level. <h3>Design</h3> Repeated-measures crossover study. <h3>Setting</h3> Rehabilitation hospital. <h3>Participants</h3> Seven individuals with iSCI completed two different sessions (AIH and SHAM) in a randomized order. <h3>Interventions</h3> AIH consisted of breathing 15 brief (∼60s) periods of low oxygen (9% O2) interspersed with 60s of normoxia (21% O2), whereas SHAM consisted of repeated exposures to normoxia. <h3>Main Outcome Measures</h3> We recorded high-density surface electromyograms from biceps and triceps brachii during maximal elbow flexion and extension torques, and decomposed these signals into individual motor unit spike trains before and 60 minutes after the interventions. <h3>Results</h3> After AIH, elbow flexion and extension torque increased by 54% (p = 0.0078; g = 0.58) and 59% (p = 0.04, g = 0.63) from baseline, respectively, whereas there was no change (i.e. < 10% change and p > 0.2 in both cases) after SHAM. Across muscles, motor unit discharge rates increased by 3.96 pulses per second (∼38%; p = 0.0015; G = 0.86) during maximal efforts, from pre to post AIH. Using generalized linear mixed models, we found that changes in motor unit discharge rates explained 54% (p = 0.0033) of the variance of changes in volitional torque. <h3>Conclusions</h3> These findings suggest that excitability and/or the activation of spinal motoneurons are augmented after AIH, providing a mechanism to explain AIH-induced increases in voluntary strength. Pending validation, AIH could potentially be used in conjunction with other therapies to enhance rehabilitation outcomes due to these enhancements in motor unit function and strength. <h3>Author(s) Disclosures</h3> The authors have nothing to disclose.

Improved Assessment of Muscle Excitation From Surface Electromyograms in Isometric Muscle Contractions.

We introduce two novel methods for the estimation of muscle excitation from surface electromyograms (EMGs), the so called cumulative motor unit activity index (CAI) and robust CAI (rCAI). Both methods aim to remove the detected motor unit action potential (MUAP) contributions from EMG but do not assess the individual motor unit spike trains. Instead, they directly estimate the cumulative motor unit spike train (CST). We compared the methods with the spatially averaged root-mean-square (RMS) envelope of the EMG signals and with the CST, estimated by the previously introduced convolution kernel compensation (CKC) method. The tests on synthetic EMG with known muscle excitation profiles demonstrated superior accuracy of newly introduced methods. In the case of 64 EMG channels and 20-dB noise, the RMS, CAI, rCAI, and CKC estimators, calculated on 0.125-s-long signal epochs, yielded the normalized RMS error (NRMSE) of 14.5% ± 2.8%, 4.4% ± 3.2%, 4.1% ± 1.8%, and 6.3% ± 4.6%, respectively. In the experimental signals from wrist extensors and flexors, the RMS, CAI, rCAI, and CKC estimations were compared to exerted muscle force. When calculated on 0.125-s-long signal epochs, they yielded the NRMSE of 11.2% ± 3.5%, 8% ± 5.6%, 10.7% ± 6.8%, and 9.0% ± 4.9%, respectively. Therefore, the newly introduced methods exhibit accuracy that is comparable to at least 200-times slower CKC method.

Electrographic spikes are common in wildtype mice

High-voltage rhythmic electroencephalographic (EEG) spikes have been recorded in wildtype (WT) rats during periods of light slow-wave sleep and passive wakefulness. The source of this activity is unclear but has been attributed to either an inherent form of absence epilepsy or a normal feature of rodent sleep EEG. In contrast, little is known about epileptiform spikes in WT mice. We thus characterize and quantify epileptiform discharges in WT mice for the first time. Thirty-six male WT C57 mice with 24-h wireless telemetry video-EEG recordings were manually scored by blinded reviewers to mark individual spikes and spike trains. Epileptiform spikes were detected in 100% of the recorded WT mice, and spike trains of at least three spikes were recorded in 90% of mice. The spikes were more frequent during the day than at night and were inversely correlated to each animal's locomotor activity. However, the discharges were not absent during active nighttime periods. These discharges may indicate a baseline tendency toward epileptic seizures or perhaps are benign variants of normal rodent background EEG. Nevertheless, a better understanding of baseline WT EEG activity will aid in differentiating pathological and normal EEG activity in mouse epilepsy models.

Sparse Estimation for the Assessment of Muscular Activity based on sEMG Measurements

Surface Electromyography (sEMG) is an established technology in measuring the electrical activity produced by skeletal muscles. For the purpose of medical diagnosis and automatic control of medical support systems, deriving a measure of muscular activity from sEMG signals is desired. This task is not trivial, since the electrical activity of the individual muscle fibers can only be measured as a superposition, after being filtered differently by the surrounding tissue. This paper presents a novel approach for the estimation of muscular activity based on estimating the composite spike train (CST), which represents the superposition of the individual spike trains of all motor units. To this end, a low-order linear state-space model is estimated from artificial CST input and sEMG output signals, generated by means of a numerical simulation of the underlying physiological processes. This model, obtained using standard system identification methods, is then utilized to estimate the CST input corresponding to real sEMG measurements. For the input estimation, a probabilistic factor graph-based algorithm is employed to perform sparse deconvolution. By enforcing sparsity, the influence of the omnipresent background noise in sEMG measurements on the estimated input is suppressed, and an input signal can be derived, which not only resembles real CST signals but also closely follows the time course of the generated muscle force. An evaluation based on measured sEMG signals of the respiratory muscles shows the practical applicability of the presented approach.

Preferential distribution of nociceptive input to motoneurons with muscle units in the cranial portion of the upper trapezius muscle.

Pain is associated with changes in the neural drive to muscles. For the upper trapezius muscle, surface electromyography (EMG) recordings have indicated that acute noxious stimulation in either the cranial or the caudal region of the muscle leads to a relative decrease in muscle activity in the cranial region. It is, however, not known if this adaption reflects different recruitment thresholds of the upper trapezius motor units in the cranial and caudal region or a nonuniform nociceptive input to the motor units of both regions. This study investigated these potential mechanisms by direct motor unit identification. Motor unit activity was investigated with high-density surface EMG signals recorded from the upper trapezius muscle of 12 healthy volunteers during baseline, control (intramuscular injection of isotonic saline), and painful (hypertonic saline) conditions. The EMG was decomposed into individual motor unit spike trains. Motor unit discharge rates decreased significantly from control to pain conditions by 4.0 ± 3.6 pulses/s (pps) in the cranial region but not in the caudal region (1.4 ± 2.8 pps; not significant). These changes were compatible with variations in the synaptic input to the motoneurons of the two regions. These adjustments were observed, irrespective of the location of noxious stimulation. These results strongly indicate that the nociceptive synaptic input is distributed in a nonuniform way across regions of the upper trapezius muscle.

Motor Unit Characteristics after Targeted Muscle Reinnervation.

Targeted muscle reinnervation (TMR) is a surgical procedure used to redirect nerves originally controlling muscles of the amputated limb into remaining muscles above the amputation, to treat phantom limb pain and facilitate prosthetic control. While this procedure effectively establishes robust prosthetic control, there is little knowledge on the behavior and characteristics of the reinnervated motor units. In this study we compared the m. pectoralis of five TMR patients to nine able-bodied controls with respect to motor unit action potential (MUAP) characteristics. We recorded and decomposed high-density surface EMG signals into individual spike trains of motor unit action potentials. In the TMR patients the MUAP surface area normalized to the electrode grid surface (0.25 ± 0.17 and 0.81 ± 0.46, p < 0.001) and the MUAP duration (10.92 ± 3.89 ms and 14.03 ± 3.91 ms, p < 0.01) were smaller for the TMR group than for the controls. The mean MUAP amplitude (0.19 ± 0.11 mV and 0.14 ± 0.06 mV, p = 0.07) was not significantly different between the two groups. Finally, we observed that MUAP surface representation in TMR generally overlapped, and the surface occupied by motor units corresponding to only one motor task was on average smaller than 12% of the electrode surface. These results suggest that smaller MUAP surface areas in TMR patients do not necessarily facilitate prosthetic control due to a high degree of overlap between these areas, and a neural information—based control could lead to improved performance. Based on the results we also infer that the size of the motor units after reinnervation is influenced by the size of the innervating motor neuron.

Encoding and decoding in parietal cortex during sensorimotor decision-making.

The lateral intraparietal area (LIP) of macaques has been asserted to play a fundamental role in sensorimotor decision-making. Here we dissect the neural code in LIP at the level of individual trial spike trains using a statistical approach based on generalized linear models. We show that LIP responses reflect a combination of temporally-overlapping task and decision-related signals. Our model accounts for the detailed statistics of LIP spike trains, and accurately predicts spike trains from task events on single trials. Moreover, we derive an optimal decoder for heterogeneous, multiplexed LIP responses that could be implemented in biologically plausible circuits. In contrast to interpretations of LIP as providing an instantaneous code for decision variables, we show that optimal decoding requires integrating LIP spikes over two timescales. These analyses provide a detailed understanding of the neural code in LIP, and a framework for studying the coding of multiplexed signals in higher brain areas.

Noninvasive, Accurate Assessment of the Behavior of Representative Populations of Motor Units in Targeted Reinnervated Muscles

Targeted muscle reinnervation (TMR) redirects nerves that have lost their target, due to amputation, to remaining muscles in the region of the stump with the intent of establishing intuitive myosignals to control a complex prosthetic device. In order to directly recover the neural code underlying an attempted limb movement, in this paper, we present the decomposition of high-density surface electromyographic (EMG) signals detected from three TMR patients into the individual motor unit spike trains. The aim was to prove, for the first time, the feasibility of decoding the neural drive that would reach muscles of the missing limb in TMR patients, to show the accuracy of the decoding, and to demonstrate the representativeness of the pool of extracted motor units. Six to seven flexible EMG electrode grids of 64 electrodes each were mounted over the reinnervated muscles of each patient, resulting in up to 448 EMG signals. The subjects were asked to attempt elbow extension and flexion, hand open and close, wrist extension and flexion, wrist pronation and supination, of their missing limb. The EMG signals were decomposed using the Convolution Kernel Compensation technique and the decomposition accuracy was evaluated with a signal-based index of accuracy, called pulse-to-noise ratio (PNR). The results showed that the spike trains of 3 to 27 motor units could be identified for each task, with a sensitivity of the decomposition > 90%, as revealed by PNR. The motor unit discharge rates were within physiological values of normally innervated muscles. Moreover, the detected motor units showed a high degree of common drive so that the set of extracted units per task was representative of the behavior of the population of active units. The results open a path for a new generation of human-machine interfaces in which the control signals are extracted from noninvasive recordings and the obtained neural information is based directly on the spike trains of motor neurons.

The Extraction of Neural Information from the Surface EMG for the Control of Upper-Limb Prostheses: Emerging Avenues and Challenges

Despite not recording directly from neural cells, the surface electromyogram (EMG) signal contains information on the neural drive to muscles, i.e., the spike trains of motor neurons. Using this property, myoelectric control consists of the recording of EMG signals for extracting control signals to command external devices, such as hand prostheses. In commercial control systems, the intensity of muscle activity is extracted from the EMG and used for single degrees of freedom activation (direct control). Over the past 60 years, academic research has progressed to more sophisticated approaches but, surprisingly, none of these academic achievements has been implemented in commercial systems so far. We provide an overview of both commercial and academic myoelectric control systems and we analyze their performance with respect to the characteristics of the ideal myocontroller. Classic and relatively novel academic methods are described, including techniques for simultaneous and proportional control of multiple degrees of freedom and the use of individual motor neuron spike trains for direct control. The conclusion is that the gap between industry and academia is due to the relatively small functional improvement in daily situations that academic systems offer, despite the promising laboratory results, at the expense of a substantial reduction in robustness. None of the systems so far proposed in the literature fulfills all the important criteria needed for widespread acceptance by the patients, i.e. intuitive, closed-loop, adaptive, and robust real-time ( 200 ms delay) control, minimal number of recording electrodes with low sensitivity to repositioning, minimal training, limited complexity and low consumption. Nonetheless, in recent years, important efforts have been invested in matching these criteria, with relevant steps forwards.

Dynamics and lifetime of persistent activity states in random networks of spiking neurons with strong synapses

The selective persistent activation of populations of neurons in the prefrontal cortex during active memory tasks is one of the best-studied neural correlates of a higher cognitive function, the so-called working memory. Yet, the respective roles of network and neuron properties in this activity are not fully understood. Spiking neuron network models often rely on a selective excitatory connectivity and generate spike trains that are more regular than what is experimentally observed, thus necessitating additional assumptions, such as intrinsic cellular bistability or additional external noisy input. Recently, it was shown that very simple random networks of excitatory and inhibitory spiking neurons can generate persistent asynchronous-irregular activity only by admitting a certain fraction of comparably strong synaptic weights [1,2]. A large number of experiments showed that synaptic weight distributions are indeed commonly characterized by a large fraction of weak weights and a heavy tail of quite strong weights [e.g. 2,3]. Here, we demonstrate how including such strong synaptic couplings leads to the occurrence of a firing-rate attractor at moderate rates with highly irregular individual spike trains. Based on the firing-rate transfer of the network, we show that the firing-rate dynamics becomes bistable if the synapses are sufficiently strong: in addition to the quiescent state, a second stable fixed point at moderate firing rates emerges by a saddle-node bifurcation. However, the population firing-rate is characterized by pronounced inherent fluctuations that perpetually perturb this fixed point. It is the trade-off between the size of the population fluctuations and the size of the basin of attraction that thus determines the onset and lifetime of persistent activity states. Moreover, individual neuronal activity turns out to be very irregular, switching between long periods of low firing rate to short burst-like states. We show that this is an effect of the strong coupling strength in the network combined with the finite memory time constant of the neurons. Thus, such irregular neuron dynamics can be a pure network phenomenon, and do not require intracellular bistability or additional high-variability noise as previously suggested.

Estimating individual firing frequencies in a multiple spike train record

Neuronal activity of several neurons is commonly recorded by a single electrode and then the individual spike trains are separated. If the separation is difficult or fails, then as a minimal result of the experiment, the individual firing rates are of interest. The proposed method solves the problem of their identification. This is possible under the condition that the recorded neurons are independent in their activities. The number of the neurons in the multi-unit record needs to be given (known or assumed) prior the calculation. The proposed method is based on the presence of the refractory period in neuronal firing, however, its precise value is not required. In addition to the determination of the individual firing rates the method can be used for an inference about the refractory period itself.

Determination of firing frequencies from multiple spike train record

Determination of firing frequencies from multiple spike train record

Dejittering of neural responses by use of their metric properties

We have adapted the dejittering technique [1,2] to analyze neural response patterns. In the original manuscript [2], we developed the idea of transformation-invariant stimulus processing in order to characterize local stimulus transformations that leave the neural response invariant. Currently, the method assumes that the neural responses are identical to within a fixed temporal precision, and the stimuli that induce them are variable. In this submission, we invert the dejittering method to operate on sets of variable neural responses associated with a fixed stimulus. We now ask the question “What are the natural transformations of neural responses that leave the neural message invariant?” The novelty and difficulty in this adaptation is the nature of the response set, which cannot be embedded naturally in a linear vector space. For example, a naive embedding of spike/no-spike to 0/1 in a time bin does not technically comprise a vector space, since linear combinations of such objects do not yield spike trains. Convolving spike trains with Gaussians suffers from the same constraints, while introducing additional assumptions about the metric and geometric structure of spike trains. Thus, the cost function used in stimulus dejittering (the Mahalanobis distance between a given stimulus and a Gaussian model of the set of stimuli), cannot be evaluated naturally on sets of spike trains. However, Victor and Purpura’s metric space distance (Dq) [3], does provide an intrinsic measure of the natural cost of transformation and can replace the Mahalanobis distance in the adapted algorithm. The metric space approach is based on providing a set of transformations that can convert one response into another. The distance between responses is the total cost of the “cheapest” set of transformations that can interconvert the signals. There have been several extension of the metric space approach, but, in the original, which we use, these transformations are the insertion or deletion of a spike (cost 1), or the translation of a spike in time (cost q per sample point of translation). The set of Dq s provides a measure of the variance of a response class without the need to embed the spike trains in a space where subtraction is defined. A response class is characterized by the set of within condition distances, ⟨Dq (i,j)⟩j, the average distance of spike train i to the rest of the responses to the same stimulus. With this measure of variance, we investigate the effects of temporal shifts of individual spike trains on the overall variance of the response class. As in [2], we then minimize the joint variance ⟨Dq(i,j)⟩t + σt2, where σt2 is the variance of temporal shifts and ⟨Dq(i,j)⟩t is the average distance between spike trains from the same response class after temporal shifts have been applied to them. Furthermore, a measurement of Dq corresponds to a set of modifications to a spike train, each of which occurs at a specific time. Thus, we can pinpoint specific regions of neural responses that are responsible for any observed large-scale temporal changes.

A General Likelihood Framework for Characterizing the Time Course of Neural Activity

We develop a general likelihood-based framework for use in the estimation of neural firing rates, which is designed to choose the temporal smoothing parameters that maximize the likelihood of missing data. This general framework is algorithm-independent and thus can be applied to a multitude of established methods for firing rate or conditional intensity estimation. As a simple example of the use of the general framework, we apply it to the peristimulus time histogram and kernel smoother, the methods most widely used for firing rate estimation in the electrophysiological literature and practice. In doing so, we illustrate how the use of the framework can employ the general point process likelihood as a principled cost function and can provide substantial improvements in estimation accuracy for even the most basic of rate estimation algorithms. In particular, the resultant kernel smoother is simple to implement, efficient to compute, and can accurately determine the bandwidth of a given rate process from individual spike trains. We perform a simulation study to illustrate how the likelihood framework enables the kernel smoother to pick the bandwidth parameter that best predicts missing data, and we show applications to real experimental spike train data. Additionally, we discuss how the general likelihood framework may be used in conjunction with more sophisticated methods for firing rate and conditional intensity estimation and suggest possible applications.

Emergent oscillations in networks of stochastic spiking neurons.

Networks of neurons produce diverse patterns of oscillations, arising from the network's global properties, the propensity of individual neurons to oscillate, or a mixture of the two. Here we describe noisy limit cycles and quasi-cycles, two related mechanisms underlying emergent oscillations in neuronal networks whose individual components, stochastic spiking neurons, do not themselves oscillate. Both mechanisms are shown to produce gamma band oscillations at the population level while individual neurons fire at a rate much lower than the population frequency. Spike trains in a network undergoing noisy limit cycles display a preferred period which is not found in the case of quasi-cycles, due to the even faster decay of phase information in quasi-cycles. These oscillations persist in sparsely connected networks, and variation of the network's connectivity results in variation of the oscillation frequency. A network of such neurons behaves as a stochastic perturbation of the deterministic Wilson-Cowan equations, and the network undergoes noisy limit cycles or quasi-cycles depending on whether these have limit cycles or a weakly stable focus. These mechanisms provide a new perspective on the emergence of rhythmic firing in neural networks, showing the coexistence of population-level oscillations with very irregular individual spike trains in a simple and general framework.

Towards statistical summaries of spike train data

AbstractStatistical inference has an important role in analysis of neural spike trains. While current approaches are mostly model-based, and designed for capturing the temporal evolution of the underlying stochastic processes, we focus on a data-driven approach where statistics are defined and computed in function spaces where individual spike trains are viewed as points. The first contribution of this paper is to endow spike train space with a parameterized family of metrics that takes into account different time warpings and generalizes several currently used metrics. These metrics are essentially penalized Lp norms, involving appropriate functions of spike trains, with penalties associated with time-warpings. The second contribution of this paper is to derive a notion of a mean spike train in the case when p=2. We present an efficient recursive algorithm, termed Matching-Minimization algorithm, to compute the sample mean of a set of spike trains. The proposed metrics as well as the mean computations are demonstrated using an experimental recording from the motor cortex.

Spatial and temporal correlations of spike trains in frog retinal ganglion cells

For a neuron, firing activity can be in synchrony with that of others, which results in spatial correlation; on the other hand, spike events within each individual spike train may also correlate with each other, which results in temporal correlation. In order to investigate the relationship between these two phenomena, population neurons' activities of frog retinal ganglion cells in response to binary pseudo-random checker-board flickering were recorded via a multi-electrode recording system. The spatial correlation index (SCI) and temporal correlation index (TCI) were calculated for the investigated neurons. Statistical results showed that, for a single neuron, the SCI and TCI values were highly related--a neuron with a high SCI value generally had a high TCI value, and these two indices were both associated with burst activities in spike train of the investigated neuron. These results may suggest that spatial and temporal correlations of single neuron's spiking activities could be mutually modulated; and that burst activities could play a role in the modulation. We also applied models to test the contribution of spatial and temporal correlations for visual information processing. We show that a model considering spatial and temporal correlations could predict spikes more accurately than a model does not include any correlation.

Bistable, Irregular Firing and Population Oscillations in a Modular Attractor Memory Network

Attractor neural networks are thought to underlie working memory functions in the cerebral cortex. Several such models have been proposed that successfully reproduce firing properties of neurons recorded from monkeys performing working memory tasks. However, the regular temporal structure of spike trains in these models is often incompatible with experimental data. Here, we show that the in vivo observations of bistable activity with irregular firing at the single cell level can be achieved in a large-scale network model with a modular structure in terms of several connected hypercolumns. Despite high irregularity of individual spike trains, the model shows population oscillations in the beta and gamma band in ground and active states, respectively. Irregular firing typically emerges in a high-conductance regime of balanced excitation and inhibition. Population oscillations can produce such a regime, but in previous models only a non-coding ground state was oscillatory. Due to the modular structure of our network, the oscillatory and irregular firing was maintained also in the active state without fine-tuning. Our model provides a novel mechanistic view of how irregular firing emerges in cortical populations as they go from beta to gamma oscillations during memory retrieval.