Neural network models of velocity storage in the horizontal vestibulo-ocular reflex

The vestibulo-ocular reflex (VOR) produces compensatory eye movements by utilizing head rotational velocity signals from the semicircular canals to control contractions of the extraocular muscles. In mammals, the time course of horizontal VOR is longer than that of the canal signals driving it, revealing the presence of a central integrator known as velocity storage. Although the neurons mediating VOR have been described neurophysiologically, their properties, and the mechanism of velocity storage itself, remain unexplained. Recent models of integration in VOR are based on systems of linear elements, interconnected in arbitrary ways. The present study extends this work by modeling horizontal VOR as a learning network composed of nonlinear model neurons. Network architectures are based on the VOR arc (canal afferents, vestibular nucleus (VN) neurons and extraocular motoneurons) and have both forward and lateral connections. The networks learn to produce velocity storage integration by forming lateral (commissural) inhibitory feedback loops between VN neurons. These loops overlap and interact in a complex way, forming both fast and slow VN pathways. The networks exhibit some of the nonlinear properties of the actual VOR, such as dependency of decay rate and phase lag upon input magnitude, and skewing of the response to higher magnitude sinusoidal inputs. Model VN neurons resemble their real counterparts. Both have increased time constant and gain, and decreased spontaneous rate as compared to canal afferents. Also, both model and real VN neurons exhibit rectification and skew. The results suggest that lateral inhibitory interactions produce velocity storage and also determine the properties of neurons mediating VOR. The neural network models demonstrate how commissural inhibition may be organized along the VOR pathway.     

Calibration of the head direction network: a role for symmetric angular head velocity cells

Continuous attractor networks require calibration. Computational models of the head direction (HD) system of the rat usually assume that the connections that maintain HD neuron activity are pre-wired and static. Ongoing activity in these models relies on precise continuous attractor dynamics. It is currently unknown how such connections could be so precisely wired, and how accurate calibration is maintained in the face of ongoing noise and perturbation. Our adaptive attractor model of the HD system that uses symmetric angular head velocity (AHV) cells as a training signal shows that the HD system can learn to support stable firing patterns from poorly-performing, unstable starting conditions. The proposed calibration mechanism suggests a requirement for symmetric AHV cells, the existence of which has previously been unexplained, and predicts that symmetric and asymmetric AHV cells should be distinctly different (in morphology, synaptic targets and/or methods of action on postsynaptic HD cells) due to their distinctly different functions.     

Path integration of head direction: updating a packet of neural activity at the correct speed using neuronal time constants

A key question in understanding the neural basis of path integration is how individual, spatially responsive, neurons may self-organize into networks that can, through learning, integrate velocity signals to update a continuous representation of location within an environment. It is of vital importance that this internal representation of position is updated at the correct speed, and in real time, to accurately reflect the motion of the animal. In this article, we present a biologically plausible model of velocity path integration of head direction that can solve this problem using neuronal time constants to effect natural time delays, over which associations can be learned through associative Hebbian learning rules. The model comprises a linked continuous attractor network and competitive network. In simulation, we show that the same model is able to learn two different speeds of rotation when implemented with two different values for the time constant, and without the need to alter any other model parameters. The proposed model could be extended to path integration of place in the environment, and path integration of spatial view.     

Accurate Path Integration in Continuous Attractor Network Models of Grid Cells

Grid cells in the rat entorhinal cortex display strikingly regular firing responses to the animal''s position in 2-D space and have been hypothesized to form the neural substrate for dead-reckoning. However, errors accumulate rapidly when velocity inputs are integrated in existing models of grid cell activity. To produce grid-cell-like responses, these models would require frequent resets triggered by external sensory cues. Such inadequacies, shared by various models, cast doubt on the dead-reckoning potential of the grid cell system. Here we focus on the question of accurate path integration, specifically in continuous attractor models of grid cell activity. We show, in contrast to previous models, that continuous attractor models can generate regular triangular grid responses, based on inputs that encode only the rat''s velocity and heading direction. We consider the role of the network boundary in the integration performance of the network and show that both periodic and aperiodic networks are capable of accurate path integration, despite important differences in their attractor manifolds. We quantify the rate at which errors in the velocity integration accumulate as a function of network size and intrinsic noise within the network. With a plausible range of parameters and the inclusion of spike variability, our model networks can accurately integrate velocity inputs over a maximum of ∼10–100 meters and ∼1–10 minutes. These findings form a proof-of-concept that continuous attractor dynamics may underlie velocity integration in the dorsolateral medial entorhinal cortex. The simulations also generate pertinent upper bounds on the accuracy of integration that may be achieved by continuous attractor dynamics in the grid cell network. We suggest experiments to test the continuous attractor model and differentiate it from models in which single cells establish their responses independently of each other.     

Path Integration of Head Direction: Updating a Packet of Neural Activity at the Correct Speed Using Axonal Conduction Delays

The head direction cell system is capable of accurately updating its current representation of head direction in the absence of visual input. This is known as the path integration of head direction. An important question is how the head direction cell system learns to perform accurate path integration of head direction. In this paper we propose a model of velocity path integration of head direction in which the natural time delay of axonal transmission between a linked continuous attractor network and competitive network acts as a timing mechanism to facilitate the correct speed of path integration. The model effectively learns a “look-up” table for the correct speed of path integration. In simulation, we show that the model is able to successfully learn two different speeds of path integration across two different axonal conduction delays, and without the need to alter any other model parameters. An implication of this model is that, by learning look-up tables for each speed of path integration, the model should exhibit a degree of robustness to damage. In simulations, we show that the speed of path integration is not significantly affected by degrading the network through removing a proportion of the cells that signal rotational velocity.     

A Three-Threshold Learning Rule Approaches the Maximal Capacity of Recurrent Neural Networks

Understanding the theoretical foundations of how memories are encoded and retrieved in neural populations is a central challenge in neuroscience. A popular theoretical scenario for modeling memory function is the attractor neural network scenario, whose prototype is the Hopfield model. The model simplicity and the locality of the synaptic update rules come at the cost of a poor storage capacity, compared with the capacity achieved with perceptron learning algorithms. Here, by transforming the perceptron learning rule, we present an online learning rule for a recurrent neural network that achieves near-maximal storage capacity without an explicit supervisory error signal, relying only upon locally accessible information. The fully-connected network consists of excitatory binary neurons with plastic recurrent connections and non-plastic inhibitory feedback stabilizing the network dynamics; the memory patterns to be memorized are presented online as strong afferent currents, producing a bimodal distribution for the neuron synaptic inputs. Synapses corresponding to active inputs are modified as a function of the value of the local fields with respect to three thresholds. Above the highest threshold, and below the lowest threshold, no plasticity occurs. In between these two thresholds, potentiation/depression occurs when the local field is above/below an intermediate threshold. We simulated and analyzed a network of binary neurons implementing this rule and measured its storage capacity for different sizes of the basins of attraction. The storage capacity obtained through numerical simulations is shown to be close to the value predicted by analytical calculations. We also measured the dependence of capacity on the strength of external inputs. Finally, we quantified the statistics of the resulting synaptic connectivity matrix, and found that both the fraction of zero weight synapses and the degree of symmetry of the weight matrix increase with the number of stored patterns.     

Recurrent networks with short term synaptic depression

Cortical circuitry shows an abundance of recurrent connections. A widely used model that relies on recurrence is the ring attractor network, which has been used to describe phenomena as diverse as working memory, visual processing and head direction cells. Commonly, the synapses in these models are static. Here, we examine the behaviour of ring attractor networks when the recurrent connections are subject to short term synaptic depression, as observed in many brain regions. We find that in the presence of a uniform background current, the network activity can be in either of three states: a stationary attractor state, a uniform state, or a rotating attractor state. The rotation speed can be adjusted over a large range by changing the background current, opening the possibility to use the network as a variable frequency oscillator or pattern generator. Finally, using simulations we extend the network to two-dimensional fields and find a rich range of possible behaviours.     

Three-dimensional head angular velocity detection from otolith afferent signals

Afferent signals from the otolith organs can produce compensatory eye position and velocity signals which has been described as linear vestibulo-ocular reflex (LVOR). The afferent otolith signals carry information about head orientation and changes of head orientation relative to gravity. A head orientation (tilt) related position signal can be obtained from population vector coding of tonic otolith afferent signals during static or dynamic head tilts, which in turn could produce compensatory eye position signals in the LVOR. On the other hand, eye angular velocity signals may be extracted, as proposed in this study, from the population response of tilt-velocity sensitive otolith afferents. Such afferents are shown to encode instantaneous head orientation relative to gravity at onset of a head movement and, as the movement continues, the projection of head angular velocity onto the earth-horizontal plane, indicating the instantaneous direction of movement relative to gravity. Angular velocity components along the earth-vertical direction which are not directly encoded by otolith afferents can be detected by central signal processing. Central reconstruction of 3D head angular velocity allows to obtain information about absolute head orientation in space even in the absence of semicircular canal related information. Such information is important for generating compensatory eye movements as well as for dynamic control of posture.     

Cell Assembly Dynamics of Sparsely-Connected Inhibitory Networks: A Simple Model for the Collective Activity of Striatal Projection Neurons

Striatal projection neurons form a sparsely-connected inhibitory network, and this arrangement may be essential for the appropriate temporal organization of behavior. Here we show that a simplified, sparse inhibitory network of Leaky-Integrate-and-Fire neurons can reproduce some key features of striatal population activity, as observed in brain slices. In particular we develop a new metric to determine the conditions under which sparse inhibitory networks form anti-correlated cell assemblies with time-varying activity of individual cells. We find that under these conditions the network displays an input-specific sequence of cell assembly switching, that effectively discriminates similar inputs. Our results support the proposal that GABAergic connections between striatal projection neurons allow stimulus-selective, temporally-extended sequential activation of cell assemblies. Furthermore, we help to show how altered intrastriatal GABAergic signaling may produce aberrant network-level information processing in disorders such as Parkinson’s and Huntington’s diseases.     

Correlations and Functional Connections in a Population of Grid Cells

We study the statistics of spike trains of simultaneously recorded grid cells in freely behaving rats. We evaluate pairwise correlations between these cells and, using a maximum entropy kinetic pairwise model (kinetic Ising model), study their functional connectivity. Even when we account for the covariations in firing rates due to overlapping fields, both the pairwise correlations and functional connections decay as a function of the shortest distance between the vertices of the spatial firing pattern of pairs of grid cells, i.e. their phase difference. They take positive values between cells with nearby phases and approach zero or negative values for larger phase differences. We find similar results also when, in addition to correlations due to overlapping fields, we account for correlations due to theta oscillations and head directional inputs. The inferred connections between neurons in the same module and those from different modules can be both negative and positive, with a mean close to zero, but with the strongest inferred connections found between cells of the same module. Taken together, our results suggest that grid cells in the same module do indeed form a local network of interconnected neurons with a functional connectivity that supports a role for attractor dynamics in the generation of grid pattern.     

Network Models of Frequency Modulated Sweep Detection

Frequency modulated (FM) sweeps are common in species-specific vocalizations, including human speech. Auditory neurons selective for the direction and rate of frequency change in FM sweeps are present across species, but the synaptic mechanisms underlying such selectivity are only beginning to be understood. Even less is known about mechanisms of experience-dependent changes in FM sweep selectivity. We present three network models of synaptic mechanisms of FM sweep direction and rate selectivity that explains experimental data: (1) The ‘facilitation’ model contains frequency selective cells operating as coincidence detectors, summing up multiple excitatory inputs with different time delays. (2) The ‘duration tuned’ model depends on interactions between delayed excitation and early inhibition. The strength of delayed excitation determines the preferred duration. Inhibitory rebound can reinforce the delayed excitation. (3) The ‘inhibitory sideband’ model uses frequency selective inputs to a network of excitatory and inhibitory cells. The strength and asymmetry of these connections results in neurons responsive to sweeps in a single direction of sufficient sweep rate. Variations of these properties, can explain the diversity of rate-dependent direction selectivity seen across species. We show that the inhibitory sideband model can be trained using spike timing dependent plasticity (STDP) to develop direction selectivity from a non-selective network. These models provide a means to compare the proposed synaptic and spectrotemporal mechanisms of FM sweep processing and can be utilized to explore cellular mechanisms underlying experience- or training-dependent changes in spectrotemporal processing across animal models. Given the analogy between FM sweeps and visual motion, these models can serve a broader function in studying stimulus movement across sensory epithelia.     

Recurrent circuitry dynamically shapes the activation of piriform cortex

In the piriform cortex, individual odorants activate a unique ensemble of neurons that are distributed without discernable spatial order. Piriform neurons receive convergent excitatory inputs from random collections of olfactory bulb glomeruli. Pyramidal cells also make extensive recurrent connections with other excitatory and inhibitory neurons. We introduced channelrhodopsin into the piriform cortex to characterize these intrinsic circuits and to examine their contribution to activity driven by afferent bulbar inputs. We demonstrated that individual pyramidal cells are sparsely interconnected by thousands of excitatory synaptic connections that extend, largely undiminished, across the piriform cortex, forming a large excitatory network that can dominate the bulbar input. Pyramidal cells also activate inhibitory interneurons that mediate strong, local feedback inhibition that scales with excitation. This recurrent network can enhance or suppress bulbar input, depending on whether the input arrives before or after the cortex is activated. This circuitry may shape the ensembles of piriform cells that encode odorant identity.     

Robust spatial working memory through homeostatic synaptic scaling in heterogeneous cortical networks

The concept of bell-shaped persistent neural activity represents a cornerstone of the theory for the internal representation of analog quantities, such as spatial location or head direction. Previous models, however, relied on the unrealistic assumption of network homogeneity. We investigate this issue in a network model where fine tuning of parameters is destroyed by heterogeneities in cellular and synaptic properties. Heterogeneities result in the loss of stored spatial information in a few seconds. Accurate encoding is recovered when a homeostatic mechanism scales the excitatory synapses to each cell to compensate for the heterogeneity in cellular excitability and synaptic inputs. Moreover, the more realistic model produces a wide diversity of tuning curves, as commonly observed in recordings from prefrontal neurons. We conclude that recurrent attractor networks in conjunction with appropriate homeostatic mechanisms provide a robust, biologically plausible theoretical framework for understanding the neural circuit basis of spatial working memory.     

Emergent Oscillations in a Realistic Network: The Role of Inhibition and the Effect of the Spatiotemporal Distribution of the Input

We have simulated a network of 10,000 two-compartment cells, spatially distributed on a two-dimensional sheet; 15% of the cells were inhibitory. The input to the network was spatially delimited. Global oscillations frequently were achieved with a simple set of connectivity rules. The inhibitory neurons paced the network, whereas the excitatory neurons amplified the input, permitting oscillations at low-input intensities. Inhibitory neurons were active over a greater area than excitatory ones, forming a ring of inhibition. The oscillation frequency was modulated to some extent by the input intensity, as has been shown experimentally in the striate cortex, but predominantly by the properties of the inhibitory neurons and their connections: the membrane and synaptic time constants and the distribution of delays.In networks that showed oscillations and in those that did not, widely distributed inputs could lead to the specific recruitment of the inhibitory neurons and to near zero activity of the excitatory cells. Hence the spatial distribution of excitatory inputs could provide a means of selectively exciting or inhibiting a target network. Finally, neither the presence of oscillations nor the global spike activity provided any reliable indication of the level of excitatory output from the network.     

Nonuniformity in the linear network model of the oculomotor integrator produces approximately fractional-order dynamics and more realistic neuron behavior

The oculomotor integrator is a network that is composed of neurons in the medial vestibular nuclei and nuclei prepositus hypoglossi in the brainstem. Those neurons act approximately as fractional integrators of various orders, converting eye velocity commands into signals that are intermediate between velocity and position. The oculomotor integrator has been modeled as a network of linear neural elements, the time constants of which are lengthened by positive feedback through reciprocal inhibition. In this model, in which each neuron reciprocally inhibits its neighbors with the same Gaussian profile, all model neurons behave as identical, first-order, low-pass filters with dynamics that do not match the variable, approximately fractional-order dynamics of the neurons that compose the actual oculomotor integrator. Fractional-order integrators can be approximated by weighted sums of first-order, low-pass filters with diverse, broadly distributed time constants. Dynamic systems analysis reveals that the model integrator indeed has many broadly distributed time constants. However, only one time constant is expressed in the model due to the uniformity of its network connections. If the model network is made nonuniform by removing the reciprocal connections to and from a small number of neurons, then many more time constants are expressed. The dynamics of the neurons in the nonuniform network model are variable, approximately fractional-order, and resemble those of the neurons that compose the actual oculomotor integrator. Completely removing the connections to and from a neuron is equivalent to eliminating it, an operation done previously to demonstrate the robustness of the integrator network model. Ironically, the resulting nonuniform network model, previously supposed to represent a pathological integrator, may in fact represent a healthy integrator containing neurons with realistically variable, approximately fractional-order dynamics. Received: 8 August 1997 / Accepted in revised form: 18 June 1998     

Emergent oscillations in a realistic network: the role of inhibition and the effect of the spatiotemporal distribution of the input.

We have simulated a network of 10,000 two-compartment cells, spatially distributed on a two-dimensional sheet; 15% of the cells were inhibitory. The input to the network was spatially delimited. Global oscillations frequently were achieved with a simple set of connectivity rules. The inhibitory neurons paced the network, whereas the excitatory neurons amplified the input, permitting oscillations at low-input intensities. Inhibitory neurons were active over a greater area than excitatory ones, forming a ring of inhibition. The oscillation frequency was modulated to some extent by the input intensity, as has been shown experimentally in the striate cortex, but predominantly by the properties of the inhibitory neurons and their connections: the membrane and synaptic time constants and the distribution of delays. In networks that showed oscillations and in those that did not, widely distributed inputs could lead to the specific recruitment of the inhibitory neurons and to near zero activity of the excitatory cells. Hence the spatial distribution of excitatory inputs could provide a means of selectively exciting or inhibiting a target network. Finally, neither the presence of oscillations nor the global spike activity provided any reliable indication of the level of excitatory output from the network.     

Weighted cue integration in the rodent head direction system

How the brain combines information from different sensory modalities and of differing reliability is an important and still-unanswered question. Using the head direction (HD) system as a model, we explored the resolution of conflicts between landmarks and background cues. Sensory cue integration models predict averaging of the two cues, whereas attractor models predict capture of the signal by the dominant cue. We found that a visual landmark mostly captured the HD signal at low conflicts: however, there was an increasing propensity for the cells to integrate the cues thereafter. A large conflict presented to naive rats resulted in greater visual cue capture (less integration) than in experienced rats, revealing an effect of experience. We propose that weighted cue integration in HD cells arises from dynamic plasticity of the feed-forward inputs to the network, causing within-trial spatial redistribution of the visual inputs onto the ring. This suggests that an attractor network can implement decision processes about cue reliability using simple architecture and learning rules, thus providing a potential neural substrate for weighted cue integration.     

Activity-dependent matching of excitatory and inhibitory inputs during refinement of visual receptive fields

The receptive field (RF) of single visual neurons undergoes progressive refinement during development. It remains largely unknown how the excitatory and inhibitory inputs on single developing neurons are refined in a coordinated manner to allow the formation of functionally correct circuits. Using whole-cell voltage-clamp recording from Xenopus tectal neurons, we found that RFs determined by excitatory and inhibitory inputs in more mature tectal neurons are spatially matched, with each spot stimulus evoking balanced synaptic excitation and inhibition. This emerges during development through a gradual reduction in the RF size and a transition from disparate to matched topography of excitatory and inhibitory inputs to the tectal neurons. Altering normal spiking activity of tectal neurons by either blocking or elevating GABA(A) receptor activity significantly impeded the developmental reduction and topographic matching of RFs. Thus, appropriate inhibitory activity is essential for the coordinated refinement of excitatory and inhibitory connections.     

Computer simulation of brainstem respiratory activity.

A mathematical model of the medullary respiratory oscillator, composed of two mutually inhibiting populations (inspiratory and expiratory) of computer-simulated neurons, is presented. Each population consists of randomly interconnected subpopulations of excitatory and inhibitory neurons, is presented. Each population consists of randomly interconnected subpopulations of excitatory and inhibitory neurons. Neuronal coupling is such that either the inspiratory or expiratory population alone is capable of cyclic activity. Weak inhibitory connections between inspiratory and expiratory populations provide satisfactory reciprocating activity independent of the natural frequency of either population alone. Initiation and persistence of rhythmic activity is dependent on a diffused noncyclic excitatory input. Vagal discharge, simulated by phasic inhibition of inspiratory neurons, results in increased respiratory frequency with decreased inspiratory activity. In the absence of simulated vagal discharge, uniform facilitation of synaptic connections increases averaged activities of inspiratory and expiratory populations, with minor effect on frequency. In the presence of simulated vagal discharge, facilitation of synaptic connections increases both frequency and amplitude. The simulated effects of synaptic facilitation, with and without vagal discharge, mimic the physiological response to CO2 in the intact and vagotimized animal.     

The Correlation Structure of Local Neuronal Networks Intrinsically Results from Recurrent Dynamics

Correlated neuronal activity is a natural consequence of network connectivity and shared inputs to pairs of neurons, but the task-dependent modulation of correlations in relation to behavior also hints at a functional role. Correlations influence the gain of postsynaptic neurons, the amount of information encoded in the population activity and decoded by readout neurons, and synaptic plasticity. Further, it affects the power and spatial reach of extracellular signals like the local-field potential. A theory of correlated neuronal activity accounting for recurrent connectivity as well as fluctuating external sources is currently lacking. In particular, it is unclear how the recently found mechanism of active decorrelation by negative feedback on the population level affects the network response to externally applied correlated stimuli. Here, we present such an extension of the theory of correlations in stochastic binary networks. We show that (1) for homogeneous external input, the structure of correlations is mainly determined by the local recurrent connectivity, (2) homogeneous external inputs provide an additive, unspecific contribution to the correlations, (3) inhibitory feedback effectively decorrelates neuronal activity, even if neurons receive identical external inputs, and (4) identical synaptic input statistics to excitatory and to inhibitory cells increases intrinsically generated fluctuations and pairwise correlations. We further demonstrate how the accuracy of mean-field predictions can be improved by self-consistently including correlations. As a byproduct, we show that the cancellation of correlations between the summed inputs to pairs of neurons does not originate from the fast tracking of external input, but from the suppression of fluctuations on the population level by the local network. This suppression is a necessary constraint, but not sufficient to determine the structure of correlations; specifically, the structure observed at finite network size differs from the prediction based on perfect tracking, even though perfect tracking implies suppression of population fluctuations.