When electrodes are placed on the scalp, or lowered into the brain itself, rhythmic waves of electrical activity are seen that reflect the coordinated firing of large numbers of neurons. The pattern of the waves varies between different brain regions, and according to what the animal or person is doing. During sleep and quiet wakefulness, slower brain waves predominate, whereas faster waves called gamma oscillations emerge during cognition—the act of processing knowledge.

Gamma waves can be readily detected in a region of the brain called the medial entorhinal cortex (MEC). This brain region is also known for its role in forming the spatial memories that allow an individual to remember how to navigate around an area they have previously visited. Individual MEC cells increase their firing rates whenever an individual is at specific locations. When these locations are plotted in two dimensions, they form a hexagonal grid: this ‘grid cell map’ enables the animal to keep track of its position as it navigates through its environment.

To determine how MEC neurons can simultaneously encode spatial locations and generate the gamma waves implicated in cognition, Solanka et al. have used supercomputing to simulate the activity of more than 1.5 million connections between MEC cells. Changing the strength of these connections had different effects on the ability of the MEC to produce gamma waves or spatial maps. However, adjusting the model to include random fluctuations in neuronal firing, or ‘noise’, was beneficial for both types of output. This is partly because noise prevented neuronal firing from becoming excessively synchronized, which would otherwise have caused seizures.

Although noise is generally regarded as disruptive, the results of Solanka et al. suggest that it helps the MEC to perform its two distinct roles. Specifically, the presence of noise enables relatively small changes in the strength of the connections between neurons to alter gamma waves—and thus affect cognition—without disrupting the neurons' ability to encode spatial locations. Given that noise reduces the likelihood of seizures, the results also raise the possibility that introducing noise into the brain in a controlled way could have therapeutic benefits for individuals with epilepsy.

Cognitive processes are mediated by computations in neural circuits and are often associated with gamma frequency oscillations in circuit activity. Gamma activity and cognitive performance often co-vary within tasks and between individuals, while cognitive deficits in psychiatric disorders such as autism and schizophrenia are linked to altered gamma frequency network dynamics (Uhlhaas and Singer, 2012; Spellman and Gordon, 2014). Such disorders are also linked to changes in the efficacy of excitatory glutamatergic and inhibitory GABAergic synapses (Rubenstein and Merzenich, 2003; Lewis et al., 2012). A critical and unresolved issue is the mechanistic relationship between gamma oscillations, the strength of excitation and inhibition, and circuit computations. On the one hand, neural codes based on firing rates may be sufficient for circuit computations (Shadlen and Newsome, 1994; Histed and Maunsell, 2014). In this scenario gamma oscillations might index circuit activation, but would not be required for computation. Evidence that rate coded computations and gamma oscillations arise from shared circuit mechanisms could be interpreted to support this view (Lundqvist et al., 2010; Pastoll et al., 2013), which predicts that when synaptic properties of a circuit are altered then gamma activity and the output of the rate-coded computation will co-vary. Alternatively, gamma oscillations, while sharing cellular substrates with rate-coded computations, may nevertheless support independent or multiplexed computational modes. For example, according to the communication through coherence hypothesis, tuning of gamma frequency activity may facilitate selective interactions between distant brain regions (Fries, 2009). In this scenario independent control of rate coded computation and gamma activity would be beneficial, for example by allowing tuning of coherence without disrupting multiplexed rate-coded computations. However, it is unclear how this could be achieved in circuits where gamma and rate-coded computations share common synaptic mechanisms, as this would require variation in synaptic properties to differentially affect gamma activity and the rate coded computation.

We address these issues using a model that accounts, through a common synaptic mechanism, for gamma oscillations and spatial computation by neurons in layer 2 of the medial entorhinal cortex (MEC) (Pastoll et al., 2013). The rate-coded firing of grid cells in the MEC is a well-studied feature of neural circuits for spatial cognition (Moser and Moser, 2013). During exploration of an environment individual grid cells are active at multiple locations that together follow a hexagonal grid-like organization. At the same time MEC circuits generate periods of activity in the high gamma frequency range (60–120 Hz) nested within a slower theta (8–12 Hz) frequency network oscillation (Chrobak and Buzsáki, 1998). Analysis of spatial correlations in grid firing, of manipulations to grid circuits, and recording of grid cell membrane potential in behaving animals, collectively point towards continuous two-dimensional network attractor states as explanations for grid firing (Bonnevie et al., 2013; Domnisoru et al., 2013; Schmidt-Hieber and Häusser, 2013; Yoon et al., 2013). In layer II of the MEC, which has the highest known density of grid cells (Sargolini et al., 2006), stellate cells that project to the dentate gyrus of the hippocampus are the major population of excitatory neurons (Gatome et al., 2010). These excitatory (E) neurons do not appear to influence one another directly but instead interact via intermediate inhibitory (I) neurons (Dhillon and Jones, 2000; Couey et al., 2013; Pastoll et al., 2013). Models that explicitly incorporate this recurrent E-I-E connectivity can account for grid firing through velocity-dependent update of network attractor states (Pastoll et al., 2013). When these models are implemented with excitable spiking neurons they also account for theta-nested gamma frequency network oscillations (Pastoll et al., 2013). The influence in these, or other classes of attractor network models, of the strength of E to I or I to E connections on gamma oscillations and grid firing, or other attractor computations, has not been systematically investigated.

We find that while gamma oscillations and grid firing are both sensitive to the strength of excitatory and inhibitory connections, their relationship differs. Although their underlying synaptic substrates are identical, gamma activity nevertheless provides little information about grid firing or the presence of underlying network attractor states. Thus, gamma activity is not a good predictor of rate-coded computation. Unexpectedly, we find the range of E- and I- synaptic strengths that support gamma and grid firing is massively increased by moderate intrinsic noise through a mechanism involving suppression of seizure-like events. In the presence of moderate noise differences in synaptic strength can tune the amplitude and frequency of gamma across a wide range with little effect on grid firing. We obtain similar results in implementations of E-I models in which connectivity is probabilistic and in models extended to include additional I to I and E to E connections. Our results suggest constraints for extrapolation of differences in gamma activity to mechanisms for cognition, identify noise as a critical factor for successful circuit computation, and suggest that tuning of excitatory or inhibitory synaptic strength could be used to control gamma-dependent processes multiplexed within circuits carrying out rate coded computations.

To systematically explore relationships between strengths of excitatory and inhibitory synapses, computations and gamma activity, we initially take advantage of models that account for both grid firing and theta-nested gamma oscillations through E-I-E interactions (Pastoll et al., 2013). In these models a layer of E cells sends synaptic connections to a layer of I cells, which in turn feedback onto the E cell layer (Figure 1A). For attractor dynamics to emerge the strength of E and I connections are set to depend on the relative locations of neurons in network space (Figure 1B). While suitable connectivity could arise during development through spike timing-dependent synaptic plasticity (Widloski and Fiete, 2014), here the connection profiles are fixed (Pastoll et al., 2013). To vary the strength of excitatory or inhibitory connections in the network as a whole we scale the strength of all connections relative to a maximum conductance value (g_E or g_I for excitation and inhibition respectively) (Figure 1B). We also consider networks in which the connection probability, rather than its strength, varies according to the relative position of neurons in the network (Figure 1—figure supplement 1). Each E and I cell is implemented as an exponential integrate and fire neuron and so its membrane potential approximates the dynamics of a real neuron, as opposed to models in which synaptic input directly updates a spike rate parameter. Addition of noise to a single E or I cell increases variability in its membrane potential trajectory approximating that seen in vivo (Figure 1C) (Domnisoru et al., 2013; Pastoll et al., 2013; Schmidt-Hieber and Häusser, 2013). Given that all neurons in the model are implemented as exponential integrate-and-fire neurons and that in total the model contains >1.5 million synaptic connections, we optimized a version of the model to enable relatively fast simulation and automated extraction and analysis of generated data (see ‘Materials and methods’). In this way the effect on grid firing of 31 × 31 combinations of g_E and g_I could be evaluated typically using >50 nodes on a computer cluster in approximately 1 week.

Figure 1 with 1 supplement see all

Download asset Open asset

Intrinsic noise increases the range of synaptic strengths that support grid firing

What happens to grid firing patterns when the strengths of excitatory and/or inhibitory synaptic connections in the model are modified? To address this we first evaluated grid firing while simulating exploration within a circular environment with a network from which noise sources were absent (Figure 2A). When we reduce the strength of connections from I cells by threefold and increase the strength of connections from E cells by threefold we find that grid firing is abolished (Figure 2Ab vs Figure 2Aa). Exploring the parameter space of g_E and g_I more systematically reveals a relatively restricted region that supports grid firing (Figure 2D and Supplementary file 1A–D). Rather than the required g_I and g_E being proportional to one another, this region is shifted towards low values of g_I and high g_E. Thus, the ability of recurrently connected networks to generate grid fields requires specific tuning of synaptic connection strengths.

Figure 2 with 4 supplements see all

Download asset Open asset

Because neural activity in the brain is noisy (Shadlen and Newsome, 1994; Faisal et al., 2008), we wanted to know if the ability of the circuit to compute location is affected by noise intrinsic to each neuron (Figure 1C). Given that continuous attractor networks are often highly sensitive to noise (Zhang, 1996; Eliasmith, 2005), we expected that intrinsic noise would reduce the parameter space in which computation is successful. In contrast, when we added noise with standard deviation of 150 pA to the intrinsic dynamics of each neuron, we found that both configurations from Figure 2Aa,b now supported grid firing patterns (Figure 2Ba,b). When we considered the full space of E and I synaptic strengths in the presence of this moderate noise we now found a much larger region that supports grid firing (Figure 2E and Supplementary file 1E–H). This region has a crescent-like shape, with arms of relatively high g_I and low g_E, and low g_I and high g_E. Thus, while tuning of g_I and g_E continues to be required for grid firing, moderate noise massively increases the range of g_E and g_I over which grid fields are generated.

When intrinsic noise was increased further, to 300 pA, the parameter space that supports grid firing was reduced in line with our initial expectations (Figure 2Ca,b,F and Supplementary file 1I–L). To systematically explore the range of g_E and g_I over which the network is most sensitive to the beneficial effects of noise we subtracted grid scores for simulations with 150 pA noise from scores with deterministic simulations (Figure 2G). This revealed that the unexpected beneficial effect of noise was primarily in the region of the parameter space where recurrent inhibition was strong. In this region, increasing noise above a threshold led to high grid scores, while further increases in noise progressively impaired grid firing (Figure 2H). In probabilistically connected networks, the range of g_E and g_I supporting grid firing was reduced, but the shape of the parameter space and dependence on noise was similar to the standard networks (Figure 2—figure supplement 1), indicating that the dependence of grid firing on g_E and g_I, and the effects of noise, are independent of the detailed implementation of the E-I attractor networks.

How closely does the firing of I cells in the simulated networks correspond to inhibitory activity in behaving animals, and to what extent is the pattern of I cell firing affected by g_E, g_I and noise? While there is little data on the spatial firing of interneurons in the MEC, recent evidence indicates that the majority of parvalbumin positive interneurons have firing fields with significant spatial stability, but low spatial sparsity and grid scores compared to excitatory grid cells (Buetfering et al., 2014). A possible interpretation of these data is that parvalbumin positive cells are unlikely to fulfill the roles of I cells predicted in E-I models. However, in networks that we evaluate here in which E cells have grid firing fields in the presence of moderate noise, I cell firing fields also have a much lower spatial information content and spatial sparsity than the corresponding E cell firing fields (E cells: spatial sparsity 0.788 ± 0.061, spatial information: 1.749 ± 0.32 bits/spike; I cells: spatial sparsity 0.239 ± 0.018, spatial information 0.243 ± 0.024 bits/spike; p < 10⁻¹⁶ for comparisons of both spatial sparsity and information; paired t-test; data range is indicated as mean ± standard deviation) (Figure 2A–C and Figure 2—figure supplement 2). Spatial autocorrelograms of simulated I cell firing fields also do not contain the six hexagonally organized peaks that are characteristic of grid fields (Figure 2A–C). Nevertheless, I cell spatial autocorrelograms produce positive grid scores (0.39 ± 0.16; Figure 2—figure supplement 3), although these are reduced compared to scores for the E cells in the same networks (E cells: 0.796 ± 0.157; p < 10⁻¹⁶; paired t-test; mean ± SD) and in many networks are below the threshold considered previously to qualify as grid like (cf. Figure 4B of Buetfering et al., 2014). When we evaluated the dependence of I cell spatial firing on g_E, g_I and noise, it appeared to be similar to that of E cells (Figure 2—figure supplement 3). To assess whether grid scores of I cells can be reduced further in E-I networks while maintaining grid firing by E cells, we investigated networks in which uncorrelated spatial input is applied to each I cell (Figure 2—figure supplement 4). In these simulations E cells had grid scores of 0.57 ± 0.25, spatial sparsity of 0.78 ± 0.03 and spatial information of 1.69 ± 0.18 bits/spike, whereas I cells had grid scores of 0.16 ± 0.2 (p < 10⁻¹⁶, paired t-test), spatial sparsity of 0.21 ± 0.01 (p < 10⁻¹⁶, paired t-test) and spatial information of 0.2 ± 0.01 bits/spike (p < 10⁻¹⁶, paired t-test; range of all data sets is mean ± SD). Thus, spatial firing of I cells has a similar dependence on noise, g_E and g_I to grid cells, conventional indices of spatial firing are nevertheless much lower for I cells in E-I networks compared to E cells, and grid firing by E cells in E-I networks is relatively robust to disruption of the rotational symmetry of I cell firing fields.

Together these simulations demonstrate that attractor circuit computations that generate grid firing fields require specific tuning of g_E and g_I. In the absence of noise grid firing is supported in relatively restricted regions of parameter space. Optimal levels of noise, which produce single cell membrane potential fluctuations of a similar amplitude to experimental observations (Domnisoru et al., 2013; Pastoll et al., 2013; Schmidt-Hieber and Häusser, 2013), promote grid firing by reducing the sensitivity of grid computations to the strength of recurrent synaptic connections, particularly when inhibition is relatively strong and excitation is weak.

Differential sensitivity of gamma oscillations and grid firing to the strength of E and I synapses

Is the sensitivity of gamma frequency oscillations to synaptic strength and to noise similar to that of grid firing? To evaluate gamma activity we recorded synaptic currents from single E and I cells across multiple theta cycles (Figure 3A–C). For the network configurations illustrated in Figure 2Aa,b and in which intrinsic noise is absent, we observed synaptic events entrained to theta cycles (Figure 3Aa,b). However, the timing and amplitude of synaptic events typically differed between theta cycles and no consistent gamma rhythm was apparent. In contrast, in the presence of noise with standard deviation 150 pA we observed nested gamma frequency synaptic activity with timing that was consistent between theta cycles (Figure 3Ba). In this condition the frequency of the gamma oscillations was reduced and their amplitude increased by raising g_I and lowering g_E (Figure 3Bb). With a further increase in noise to 300 pA, gamma activity remained entrained to theta cycles, but became less ordered (Figure 3Ca,b).

Figure 3 with 4 supplements see all

Download asset Open asset

To explore gamma activity across a wider range of g_I and g_E we automated quantification of the strength and frequency of oscillatory input to E cells (see ‘Materials and methods’). In the absence of noise gamma frequency activity only occurred for a narrow range of g_I and g_E (Figure 3D). Strikingly, following addition of moderate noise the region of parameter space that supports gamma activity was massively expanded (Figure 3E). Within this space, the amplitude of gamma increased with increasing inhibition, whereas the frequency was reduced. As noise is increased further the amplitude and frequency of gamma oscillations are reduced (Figure 3F). We found a similar dependence of gamma oscillations on noise, g_E and g_I in networks with probabilistic connectivity (Figure 3—figure supplement 1). Thus, intrinsic noise modifies the amplitude and frequency of nested gamma oscillations.

To determine whether there is a systematic relationship between values of g_E and g_I that generate gamma and grid firing we compared the gridness score and gamma scores for each circuit configuration (Figure 3G, Figure 3—figure supplements 2, 3). We found this relationship to be complex and highly sensitive to noise. However, we did not find any evidence for strong linear relationships between gamma amplitude or gamma frequency and grid score (R² < 0.12 for all comparisons), while gamma amplitude and frequency provided only modest amounts of information about grid scores (0.27 < MIC < 0.33 and 0.27 < MIC < 0.37 respectively). The relationship between noise intensity and gamma differed from that for grid computations. Whereas, grids emerged above a sharp noise threshold (Figure 2H), for the same regions in parameter space the frequency and amplitude of gamma oscillations varied smoothly as a function of noise (Figure 3H). Thus, neither the frequency nor the power of gamma appears to be a good predictor of grid firing.

When we considered only regions of parameter space that generate robust grid fields (grid score >0.5), we found circuits generating almost the complete observed range of gamma amplitudes (0.02 < autocorrelation peak < 0.59) and frequencies (31 Hz < frequency < 102 Hz) (Figure 3—figure supplement 4). For example, considering the crescent shaped region of E-I space that supports grid firing in the presence of intermediate noise (the region within the isocline in Figure 3E), when g_I is high and g_E low then the amplitude of gamma is relatively low and the frequency high. Moving towards the region where g_I is high and g_E is low, the amplitude of gamma is increased and the frequency is reduced. Thus, variation of synaptic strength across this region of E-I space can be used to tune the properties of gamma activity while maintaining the ability of the network to generate grid fields.

Together these data indicate that an optimal level of noise promotes emergence of gamma oscillations, while the properties of oscillations may depend on the relative strength of synaptic connections. The relationship between gamma and synaptic strength differs to that for grid computations. Strikingly, while gamma activity provides relatively little information about grid firing, differential sensitivity of gamma and grid firing to g_E and g_I provides a mechanism for circuits to tune gamma frequency activity while maintaining the ability to compute rate coded grid firing fields.

Noise promotes attractor computation by opposing seizures

Given the emergence of a large parameter space that supports grid firing following introduction of moderate noise, we were interested to understand how noise influences the dynamics of the E-I circuits. One possibility is that in networks that fail to generate grid firing fields network attractor states form, but their activity bumps are unable to track movement. In this scenario disrupted grid firing would reflect incorrect control of network activity by velocity signals. Alternatively, deficits in grid firing may reflect failure of network attractor states to emerge. To distinguish these possibilities we investigated formation of activity bumps in network space over the first 10 s following initialization of each network (Figure 4).

Figure 4 with 1 supplement see all

Download asset Open asset

Our analysis suggests that the deficit in grid firing in deterministic compared to noisy networks reflects a failure of attractor states to emerge. For deterministic simulation of the points in parameter space considered in Figure 2Aa, which are able to generate grid patterns, we found that a single stable bump of activity emerged over the first 2.5 s of simulated time (Figure 4Aa). In contrast, for deterministic simulation of the point considered in 2Ab, which in deterministic simulations did not generate grid patterns, a single stable bump fails to emerge (Figure 4Ab). Quantification across the wider space of g_E and g_I values (see ‘Materials and methods’) indicated that when g_I is low there is a high probability of bump formation as well as grid firing, whereas when g_I is high the probability of both is reduced (Figure 4B). In contrast to the deterministic condition, for circuits with intrinsically noisy neurons activity bumps emerged in the first 1.25 s following initialization of the network (Figure 4Ac–e) and the area of parameter space that supported bump formation was much larger than that supporting grid firing (Figure 4B). Plotting gridness scores as a function of bump probability indicated that bump formation was necessary, although not sufficient for grid formation (Figure 4C), while plotting the first autocorrelation peak as a function of bump probability supported our conclusion that grid computation and gamma activity are not closely related (Figure 4D). Together, these data indicate that noise promotes formation of attractor bumps in network activity and in deterministic simulations the failure of the circuit to generate attractor states largely accounts for disrupted grid firing.

In noisy networks the presence of low grid scores for networks with high bump scores (Figure 4C) is explained by sensitivity of these network configurations to noise-induced drift. This is illustrated by the region of parameter space from Figure 2Ab, where g_I is relatively high and g_E relatively low, and which in deterministic simulations fails to generate bumps or grids. With moderate noise, this point generates bumps that show little drift (Figure 4Ac), whereas as noise is increased further the bump begins to drift (Figure 4Ae). In contrast, at the point illustrated in Figure 2Aa, which forms grids and bumps in the presence or absence of noise, activity bumps are relatively stable in each condition (Figure 4Aa,d), although drift increases with greater noise (Figure 4—figure supplement 1). Thus, intrinsic noise has two opposing effects on bump formation. For much of the parameter space we consider moderate noise promotes emergence of bumps and grids, while across all of parameter space noise reduces bump stability leading to deterioration of grids.

To investigate how addition of noise promotes emergence of network attractor states we investigated the dynamics of neurons in the simulated circuits. We focus initially on the point in parameter space identified in Figure 2Ab, where grids are found in the presence of moderate noise, and bumps are found when noise is moderate or high. When we examined times of action potentials generated by all neurons in this circuit, we find that in the absence of noise the network generates hyper-synchronous seizure-like states at the start of each theta cycle (Figure 5A and Figure 5—figure supplement 1A). The number of E cells active on each theta cycle differs, but their activity is typically restricted to the rising phase of theta, and there is no consistent structure in the pattern of activated neurons. The number of simultaneously active I cells is also greatest at the start of each theta cycle. The I-cells continue to fire over the theta cycle, but their synchronization declines. When moderate noise is added to the circuit only a subset of E-cells are active on each theta cycle, forming an activity bump (Figure 5B and Figure 5—figure supplement 1B). The I-cells are active at gamma frequency and the formation of an activity bump in the E-cell population is reflected by an inverted bump in the I-cell population activity (Figure 5B). With increased noise there is a similar overall pattern of activity, but spike timing becomes more variable, causing the bumps to drift and reducing the degree of synchronization at gamma frequencies (Figure 5C and Figure 5—figure supplement 1C).

Figure 5 with 1 supplement see all

Download asset Open asset

To determine whether these changes in network dynamics are seen across wider regions of parameter space we first quantified the presence of seizure like events from the maximum population firing rate in any 2 ms window over 10 s of simulation time (E-rate_max). Strikingly, we found that in the absence of noise epochs with highly synchronized activity were found for almost all combinations of g_E and g_I, whereas these seizure-like events were absent in simulations where noise was present (Figure 5D). Interestingly, while grids emerge in deterministic networks in regions of E-I space where E-rate_max is relatively low, there is a substantial region of parameter space in which E-rate_max is >400 Hz, but grids are nevertheless formed. It is possible that seizure-like states may be rare in this region of parameter space and so do not interfere sufficiently with attractor dynamics to prevent grid firing. To test this we calculated for each combination of g_E and g_I the proportion of theta cycles having events with population-average rate >300 Hz (P_E-rate > 300). For values of g_E and g_I where grid fields are present P_E-rate > 300 was relatively low, indicating that seizure-like events are indeed rare (Figure 5E). Consistent with this, when we plotted grid score as a function of P_E-rate > 300, we found that P_E-rate > 300 was relatively informative about the gridness score in networks without noise (MIC = 0.624) and a low value of P_E-rate > 300 was necessary for grid firing (Figure 5F). In contrast, E-rate_max was less informative of grid firing (0.392 ≤ MIC ≤0.532) and a wide range of values were consistent with grid firing (Figure 5F). Thus, while grid firing is compatible with occasional seizure-like events, when seizure-like events occur on the majority of theta cycles then grid firing is prevented.

Because seizure-like events tend to initiate early on the depolarizing phase of each theta cycle, we asked if synchronization by theta frequency drive plays a role in their initiation. When theta frequency input was replaced with a constant input with the same mean amplitude, the power of gamma oscillations was still dependent on the levels of noise and changes in g_E and g_I (Figure 6—figure supplement 1). However, in contrast to simulations with theta frequency input (Figure 5D,E), noise-free networks without theta exhibited hyper-synchronous firing only when g_E was <0.5 nS (Figure 6A) and generated grid firing fields almost in the complete range of g_E and g_I (Figure 6D,G). Addition of noise in the absence of theta had mostly detrimental effects on grid firing (Figure 6E,F,H,I and Figure 6—figure supplement 2). Interestingly, with intermediate levels of noise, the subregion with high gridness scores (>0.5) retained its crescent-like shape (Figure 6E,H), but was smaller when compared to the networks with theta frequency inputs (size of regions with and without theta: 488/961 vs 438/961), while the range of gamma frequencies present was much lower than in networks containing theta drive. Together, these data indicate that moderate noise prevents emergence of seizure like states by disrupting synchronization of the attractor network by the shared theta frequency drive. In networks with moderate noise theta drive promotes grid firing and enables a wide range of gamma frequencies to be generated without disrupting grid firing.

Figure 6 with 5 supplements see all

Download asset Open asset

Our analysis points towards suppression of seizure-like events as the mechanism by which moderate noise promotes grid firing, while interactions between noise and theta appear important for the capacity to multiplex grid firing with a wide range of gamma frequencies. However, we wanted to know if other factors might contribute to these beneficial roles of noise. Grid fields may also fail to form if overall activity levels are too low, in which case neurons with grid fields instead encode head direction (Bonnevie et al., 2013). This observation is unlikely to explain our results as the mean firing rate of E cells in networks that generated grid firing fields (grid score >0.5, networks with g_E or g_I set to 0 excluded) was in fact lower than the firing rate of networks without grid fields (1.2; 1.0; 1.0 Hz grid fields vs 3.0; 2.7; 1.2 Hz no grid fields, in networks with σ = 0; 150; 300 pA respectively). There was also no systematic relationship between grid score and firing frequency (Figure 6—figure supplement 3). We also wanted to know if other properties of grid fields vary as a function of g_E and g_I. Parameters used to calibrate velocity integration by the grid network varied very little with changes in g_E and g_I (Figure 6—figure supplement 4), whereas drift increased with g_I (Figure 4—figure supplement 1) and place cell input was most effective in opposing attractor drift in noisy networks with high gridness scores (Figure 6—figure supplement 5). These data are consistent with suppression of seizure like events as the mechanism by which noise promotes grid firing, while interactions between noise and theta frequency inputs profoundly influence the dynamics of attractor networks that generate grid fields.

Recurrent inhibition increases the frequency of gamma activity and promotes grid firing

Our analysis so far focuses on E-I attractor networks as simple models of grid firing that are compatible with the finding that synaptic interactions between stellate cells in layer 2 of the MEC are mediated via inhibitory interneurons (Dhillon and Jones, 2000; Couey et al., 2013; Pastoll et al., 2013). However, there is evidence that interneurons active during theta-nested gamma activity make connections to one another as well as to stellate cells (Pastoll et al., 2013). To establish whether this recurrent inhibition substantially modifies our conclusions from simpler E-I networks, we extended the E-I model to also include synapses between interneurons (see ‘Materials and methods’). In the resulting E-I-I networks, in the absence of noise, grid firing emerges across a much larger region of parameter space compared to E-I networks (Figure 7A, Figure 7—figure supplements 1–4). However, as in E-I networks occasional seizure like activity was present across a wide range of g_E and g_I (Figure 7—figure supplement 5), and gamma frequency activity was largely absent (Figure 7D,G). Following addition of noise with standard deviation of 150 pA to E-I-I networks, grid firing was maintained, seizure like activity was abolished, and gamma like activity emerged (Figure 7B,E,H and Figure 7—figure supplement 5). Increasing the noise amplitude to 300 pA reduced grid firing and interfered with the emergence of gamma oscillations (Figure 7C,F,I and Figure 7—figure supplements 1–5). Importantly, just as in E-I networks, the presence of moderate noise in E-I-I networks enables tuning of gamma activity by varying g_E and g_I while maintaining the ability of the networks to generate grid firing fields. Gamma activity had a higher frequency in E-I-I compared to E-I networks, with a greater proportion of the parameter space supporting gamma frequencies >80 Hz. This higher frequency gamma is similar to fast gamma observed experimentally in the MEC (cf. Chrobak and Buzsáki, 1998; Colgin et al., 2009; Pastoll et al., 2013). Thus, by including additional features of local circuits in layer 2 of the MEC, E-I-I models may more closely recapitulate experimental observations. Nevertheless, E-I-I networks maintain the ability, in the presence of moderate noise, for variation in g_E and g_I to tune gamma oscillations without interfering with grid firing.

Figure 7 with 10 supplements see all

Download asset Open asset

Finally, we asked if addition of synaptic connections between excitatory cells modifies the relationship between gamma, noise, g_E and g_I. While the E-I model is consistent with the connectivity between stellate cells in layer 2 of the MEC, adjacent pyramidal cells may also have grid firing properties. Unlike stellate cells, pyramidal cells interact with one another directly via excitatory connections and indirectly via inhibitory interneurons (Couey et al., 2013). To assess the impact of E-E connections, we first extended the E-I model to allow each E cell to excite other E cells that are nearby in neuron space. The dependence of grid firing, gamma oscillations, and bump formation on noise, g_E and g_I was similar to E-I networks (Figure 7—figure supplements 6–9). We also attempted to evaluate networks in which E-E connections were structured, but E-I and I-E connections were uniformly distributed. However, in these networks we were unable to identify parameters that support formation of stable activity bumps (Figure 7—figure supplement 10). This is consistent with instability of simpler network attractors based on E-E connections (Seung et al., 2000).

We investigated the relationship between rate coded spatial computations and nested gamma oscillations in attractor network models of grid firing. While in the models we consider rate coding and gamma oscillations share the same neural substrate, that is projections from a population of E cells to an I cell population, which in turn projects back to the E cell population, we find that their sensitivity to variations in excitatory and inhibitory synaptic strengths nevertheless differs. A moderate level of noise promotes generation of both grid fields and nested gamma oscillations, primarily by the disruption of epileptic-like firing of E and I cells in the network. When the strength of E or I connections is varied in the presence of moderate noise a wide range of gamma frequency and power can be obtained without affected grid firing. Thus, noise can be beneficial for computations performed by the nervous system, while the frequency and power of multiplexed gamma oscillations can be tuned independently of rate-coded grid computations, suggesting a mechanism for differential control of multiplexed neural codes.

Our results suggest a novel beneficial role for noise. In general noise in the nervous system is believed to distort the fidelity of transmitted signals (Faisal et al., 2008). Exceptions are stochastic resonance phenomena in which noise promotes detection of small amplitude signals by individual neurons (Longtin et al., 1991; Benzi et al., 1999; Shu et al., 2003), improvements in signal coding through desynchronization of neuronal populations (Hunsberger et al., 2014) and emergence of stochastic weak synchronization in interneuron networks (Tiesinga and Jose, 2000). The beneficial role for noise that we identify here differs from these phenomena in that it emerges through interactions between populations of neurons and because the grid cell attractor network performs a computation—generation of a spatial code from velocity inputs—rather than propagating input signals. We find that by opposing emergence of hyper-sychronous seizure-like states noise allows the network to generate stable bump attractor states. Noise prevents the seizure-like states by desynchronizing neuronal responses to common theta input. We were able to identify this role for noise because spiking and synaptic dynamics are explicitly represented in the simulated network. These dynamics are absent from other attractor network models of grid firing (Fuhs and Touretzky, 2006; Guanella et al., 2007; Burak and Fiete, 2009). They are also absent from other models of theta-nested gamma oscillations that simulate two-dimensional dynamical systems of E and I populations with theta modulated inputs to the network (Onslow et al., 2014). Thus, intrinsic cellular and synaptic dynamics in conjunction with noise sources may be important in accounting for computations and oscillatory activity in neural networks.

The distinct control of rate coded grid computations and gamma oscillations by noise, g_E and g_I was independent of the detailed implementation of the E-I models we considered and was maintained in more complex models incorporating I-I and E-E coupling. Current available experimental data appears to be insufficient to distinguish between these different models. For example, our analysis of interneuron firing indicates that while E-I models predict that interneurons will have spatial firing fields, they have lower spatial information content, spatial sparsity and grid scores than E cells and therefore may be difficult to detect in existing experimental datasets and with current analysis tools. Thus, evidence previously interpreted to argue against E-I based mechanisms for grid firing may in fact not distinguish these from other possible mechanisms. Indeed, we found that grid firing by E cells can be maintained during spatial input that distorts the spatial firing pattern of I cells (Figure 2—figure supplement 4). While these simulations establish in principle that E-I based attractor networks can generate grid outputs even when spatial firing of many E and I cells in the network is not clearly grid-like, the extent to which these networks can account for additional details of experimental observations, for example weak periodic patterns in the spatial autocorrelation of the firing fields of some PV interneurons (cf. Buetfering et al., 2014, Figure 4a), is not yet clear. Our results are consistent with local synaptic connections, in addition to those between E cells and I cells, having important functional roles. For example addition of synapses between interneurons to E-I networks causes an overall increase in the frequency of gamma activity and in the stability of grid firing. Nevertheless, we find that in these modified networks moderate noise still enables variation in g_E and g_I to tune gamma oscillations independently from grid firing.

An intriguing aspect of our results is that they suggest novel approaches to suppressing seizures and to promoting normal cognitive function. Seizures have previously been suggested to result from deficits in inhibition or from alterations in intrinsic excitability of neurons (Lerche et al., 2001; Treiman, 2001). We show that seizures can be induced when these properties are held constant simply by reducing levels of noise within a circuit. A future experimental challenge for dissecting the contribution of intrinsic noise to seizures will be to target biological noise sources. In the brain noise arises from ion channel gating and from background synaptic activity. It is therefore difficult to manipulate noise sources without also affecting intrinsic excitability or excitation-inhibition balance. However, it may be feasible to add noise to circuits through transcranial magnetic stimulation (Ruzzoli et al., 2010). In this case our simulations predict that addition of noise may restore epileptic circuits to normal activity. This mechanism may explain why focal electrical stimulation of the entorhinal cortex in patients with seizures leads to an enhancement of memory performance (Suthana et al., 2012).

While correlations between gamma oscillations and various cognitive and pathological brain states are well established, the proposed computational roles of gamma oscillations have been difficult to reconcile with rate-coded representations with which they co-exist. We were able to address this issue directly by analyzing a circuit in which gamma oscillations and rate-coded computations arise from a shared mechanism. Rather than gamma serving as an index of rate-coded computation, we find instead that there is a substantial parameter space across which rate-coded computation is stable, while the amplitude and frequency of theta-nested gamma oscillations varies. Our analysis leads to several new and testable predictions. First, tuning of recurrent synaptic connections could be used to modify gamma oscillations without affecting rate-coded computation. If multiple networks of the kind we simulate here correspond to grid modules providing input to downstream neurons in the hippocampus (Stensola et al., 2012), then adjusting g_E or g_I would alter gamma frequency with minimal effect on the grid firing pattern of each module. If the downstream neurons integrate input at the gamma time scale, then this should lead to re-mapping of their place representation in the absence of any change in either the strength of their synaptic inputs or the information they receive from upstream grid cells. Adjustment of g_E and g_I could be achieved dynamically through actions of neuromodulators (Marder, 2012), or on slower developmental time scales (Widloski and Fiete, 2014). Second, subtle differences in gamma could be a sensitive index of network pathology at stages before deficits in rate coded computation are apparent. If cognitive deficits in psychiatric disorders reflect a failure of rate coded computation, then our analysis predicts that a change in noise within a circuit, in addition to synaptic modification, may be necessary for deficits to emerge. From this perspective it is intriguing that seizure phenotypes are often associated with disorders such as autism (Deykin and MacMahon, 1979). Alternatively, cognitive deficits may result from a failure to coordinate gamma frequency synchronization of circuits that converge on downstream targets. In this case we expect cognitive deficits to be phenocopied by manipulations that affect gamma frequency or power without influencing rate-coded computations (Sigurdsson et al., 2010; Spellman and Gordon, 2014).

In conclusion, our systematic exploration of three dimensions of parameter space (g_E, g_I and intrinsic noise) illustrates the complexity of relationships between rate-coded computation, gamma frequency oscillations and underlying cellular and molecular mechanisms. Our results highlight the challenges in straightforward interpretation of experiments in which these parameters are correlated to one another, (cf. Wang and Krystal, 2014). While there are parallels to investigations of pace-making activity in invertebrate circuits (Marder and Taylor, 2011), which demonstrate that many parameter combinations can account for higher order behavior, there are also critical differences in that the models we describe account for multiplexing of rate-coded computation and oscillatory activity, while the number of neurons and connections in the simulated circuit is much larger. Future experimentation will be required to test our model predictions for unexpected beneficial roles of noise and for control of gamma oscillations independently from grid firing by modulating the strength of excitatory and inhibitory synaptic connections.

The model comprised a network of exponential integrate and fire neurons (Fourcaud-Trocmé et al., 2003) implemented as a custom-made module of the NEST simulator (Gewaltig and Diesmann, 2007). The network investigated in the majority of simulations (Figures 1–6) is modified from that in Pastoll et al. (2013) and consists of excitatory (E) and inhibitory (I) populations of neurons that were arranged on a twisted torus with dimensions of 34 × 30 neurons. In networks where connection strengths were generated probabilistically instead of in an all-to-all way, the synaptic weights from E to I cells and vice versa were constant, while the probability of connection between the pre- and post-synaptic neuron was drawn according to Figure 1B. In addition, some networks also included direct uniform recurrent inhibition between I cells (Figure 7; referred to as E-I-I networks) or direct structured recurrent excitation between E cells (Figure 7—figure supplements 6–10). When recurrent excitation was present, synaptic weights between E cells followed the connectivity profile in which the strongest connection was between cells that were close to each other in network space (Figure 1B) and the weights between E and I cells were generated either according to synaptic profiles from Figure 1B (Figure 7—figure supplements 6–9) or the E-I connectivity was uniform with a probability of connection of 0.1 (Figure 7—figure supplement 10). E and I cells also received the theta current drive which was the sum of a constant amplitude positive current and a current with sinusoidal waveform (8 Hz). The constant component of the drive was required to activate the circuit, while the theta drive frequency was chosen to reflect the frequency of theta oscillations in behaving animals. The amplitude (cf. Appendix 1) was chosen to produce theta modulation of I cell firing similar to that observed in behaving animals (cf. Chrobak and Buzsáki, 1998) and ex-vivo models of theta-nested gamma activity (cf. Pastoll et al., 2013). In order to oppose drift of the activity bump in networks that simulated exploration of the arena E cells received input from cells with place-like firing fields simulated as Poisson spiking generators with their instantaneous firing rate modeled as a Gaussian function of the animal position. Full details of the connectivity and network parameters are in Appendix 1.

In all simulations the networks were parameterized by the standard deviation of noise (σ) injected independently into each E and I cell and by synaptic scaling parameters (g_E and g_I). Noise was sampled from a Gaussian distribution with standard deviation either set to σ = 0, 150 or 300 pA, or alternatively in the range of 0–300 pA in steps of 10 pA (Figures 2H, 3H). The peaks of the synaptic profile functions (Figure 1B) were determined by the g_E and g_I parameters that appropriately scaled the maximal conductance values of the excitatory and inhibitory connections respectively.

Gridness scores were estimated by simulating exploration in a circular arena with a diameter of 180 cm. For each value of g_E and g_I the simulations consisted of two phases. In the first phase, animal movement with constant speed and direction (vertically from bottom to top) was simulated in order to calibrate the gain of the velocity input to achieve 60 cm spacing between grid fields in the network. In the second phase, the calibrated velocity input gains were used during a simulation of realistic animal movements with duration of 600 s (Hafting et al., 2005). Each simulation was repeated 1–4 times. For each trial, gridness score was then estimated from an E or I cell located at position (0, 0) on the twisted torus. In simulations where interneurons received uncorrelated spatial inputs (Figure 2—figure supplement 4), gridness scores were estimated from 100 randomly selected E and I cells on the twisted torus.

For the analysis of bump attractor properties and gamma oscillations a separate set of simulations were run. For each value of g_E, g_I and noise level, there were five trials of 10 s duration during which the velocity and place cell inputs were deactivated. For each trial spiking activity of all cells was recorded. In addition, inhibitory synaptic currents of 25 randomly selected E cells were saved and used for further analysis.

The strength and frequency of gamma oscillations were estimated from the inhibitory synaptic currents recorded from E cells. The currents were first band-pass filtered between 20 and 200 Hz. For each trace, its autocorrelation function was computed and the first local maximum was detected using a peak detection algorithm which was based on calculating the points in the autocorrelation function where the first difference of the signal changed sign from positive to negative and thus approximated the points where the first derivative was zero and the second derivative was negative. The strength and frequency of gamma oscillations was estimated from the correlation value and lag at the position of the first local maximum respectively.

Properties of bump attractors were estimated by fitting symmetric Gaussian functions onto successive snapshots of firing rates of each cell in the E population. For each snapshot this procedure yielded the position of the bump center and its width. The probability of bump formation was then estimated as a proportion of population-activity snapshots that were classified as bump attractors, that is, those fitted Gaussian functions whose width did not exceed the shorter side of the twisted torus. Other properties of bump attractors were estimated by analyzing successive positions of the bump attractor centers. Action potential raster plots of E and I populations (Figure 5A–C, Figure 5—figure supplement 1 and Figure 7—figure supplement 10) show neuron indices that are flattened in a row-wise manner with respect to the two-dimensional twisted torus. Data points with white color in Figure 5D,E and Figure 5—figure supplement 1A have been excluded from analysis since the maximal firing rate of E cells exceeded 500 Hz/2 ms window.

The calculation of the maximal information coefficient (MIC) for the relationship between gridness score, gamma and bump scores was estimated by applying the MIC measure using the minepy package (Albanese et al., 2013). Calculations of spatial information were carried out according to (Skaggs et al., 1996). Spatial sparsity was calculated by following the procedure outlined in (Buetfering et al., 2014). All other data analysis and simulations were performed in Python.

1. 1. Albanese D
   2. Filosi M
   3. Visintainer R
   4. Riccadonna S
   5. Jurman G
   6. Furlanello C

   (2013) Minerva and minepy: a C engine for the MINE suite and its R, Python and MATLAB wrappers

   Bioinformatics 29:407–408.

   https://doi.org/10.1093/bioinformatics/bts707

   • Google Scholar
2. 1. Benzi R
   2. Sutera A
   3. Vulpiani A

   (1999) The mechanism of stochastic resonance

   Journal of Physics A 14:L453–L457.

   https://doi.org/10.1088/0305-4470/14/11/006

   • Google Scholar
3. 1. Bonnevie T
   2. Dunn B
   3. Fyhn M
   4. Hafting T
   5. Derdikman D
   6. Kubie JL
   7. Roudi Y
   8. Moser EI
   9. Moser MB

   (2013) Grid cells require excitatory drive from the hippocampus

   Nature Neuroscience 16:309–317.

   https://doi.org/10.1038/nn.3311

   • Google Scholar
4. 1. Buetfering C
   2. Allen K
   3. Monyer H

   (2014) Parvalbumin interneurons provide grid cell-driven recurrent inhibition in the medial entorhinal cortex

   Nature Neuroscience 17:710–718.

   https://doi.org/10.1038/nn.3696

   • Google Scholar
5. 1. Burak Y
   2. Fiete IR

   (2009) Accurate path integration in continuous attractor network models of grid cells

   PLOS Computational Biology 5:e1000291.

   https://doi.org/10.1371/journal.pcbi.1000291

   • Google Scholar
6. 1. Chrobak JJ
   2. Buzsáki G

   (1998)

   Gamma oscillations in the entorhinal cortex of the freely behaving rat

   The Journal of Neuroscience 18:388–398.

   • Google Scholar
7. 1. Compte A
   2. Brunel N
   3. Goldman-Rakic PS
   4. Wang XJ

   (2000) Synaptic mechanisms and network dynamics underlying spatial working memory in a cortical network model

   Cerebral Cortex 10:910–923.

   https://doi.org/10.1093/cercor/10.9.910

   • Google Scholar
8. 1. Colgin LL
   2. Denninger T
   3. Fyhn M
   4. Hafting T
   5. Bonnevie T
   6. Jensen O
   7. Moser MB
   8. Moser EI

   (2009) Frequency of gamma oscillations routes flow of information in the hippocampus

   Nature 462:353–357.

   https://doi.org/10.1038/nature08573

   • Google Scholar
9. 1. Couey JJ
   2. Witoelar A
   3. Zhang SJ
   4. Zheng K
   5. Ye J
   6. Dunn B
   7. Czajkowski R
   8. Moser MB
   9. Moser EI
   10. Roudi Y
   11. Witter MP

   (2013) Recurrent inhibitory circuitry as a mechanism for grid formation

   Nature Neuroscience 16:318–324.

   https://doi.org/10.1038/nn.3310

   • Google Scholar
10. 1. Deykin EY
    2. MacMahon B

    (1979) The incidence of seizures among children with autistic symptoms

    The American Journal of Psychiatry 136:1310–1312.

    https://doi.org/10.1176/ajp.136.10.1310

    • Google Scholar
11. 1. Dhillon A
    2. Jones RS

    (2000) Laminar differences in recurrent excitatory transmission in the rat entorhinal cortex in vitro

    Neuroscience 99:413–422.

    https://doi.org/10.1016/S0306-4522(00)00225-6

    • Google Scholar
12. 1. Domnisoru C
    2. Kinkhabwala AA
    3. Tank DW

    (2013) Membrane potential dynamics of grid cells

    Nature 495:199–204.

    https://doi.org/10.1038/nature11973

    • Google Scholar
13. 1. Eliasmith C

    (2005) A unified approach to building and controlling spiking attractor networks

    Neural Computation 17:1276–1314.

    https://doi.org/10.1162/0899766053630332

    • Google Scholar
14. 1. Faisal AA
    2. Selen LP
    3. Wolpert DM

    (2008) Noise in the nervous system

    Nature Reviews. Neuroscience 9:292–303.

    https://doi.org/10.1038/nrn2258

    • Google Scholar
15. 1. Fourcaud-Trocmé N
    2. Hansel D
    3. van Vreeswijk C
    4. Brunel N

    (2003)

    How spike generation mechanisms determine the neuronal response to fluctuating inputs

    The Journal of Neuroscience 23:11628–11640.

    • Google Scholar
16. 1. Fries P

    (2009) Neuronal gamma-band synchronization as a fundamental process in cortical computation

    Annual Review of Neuroscience 32:209–224.

    https://doi.org/10.1146/annurev.neuro.051508.135603

    • Google Scholar
17. 1. Fuhs MC
    2. Touretzky DS

    (2006) A spin glass model of path integration in rat medial entorhinal cortex

    The Journal of Neuroscience 26:4266–4276.

    https://doi.org/10.1523/JNEUROSCI.4353-05.2006

    • Google Scholar
18. 1. Gatome CW
    2. Slomianka L
    3. Lipp HP
    4. Amrein I

    (2010) Number estimates of neuronal phenotypes in layer II of the medial entorhinal cortex of rat and mouse

    Neuroscience 170:156–165.

    https://doi.org/10.1016/j.neuroscience.2010.06.048

    • Google Scholar
19. 1. Gewaltig MO
    2. Diesmann M

    (2007) NEST (NEural Simulation Tool)

    Scholarpedia 2:1430.

    https://doi.org/10.4249/scholarpedia.1430

    • Google Scholar
20. 1. Guanella A
    2. Kiper D
    3. Verschure P

    (2007) A model of grid cells based on a twisted torus topology

    International Journal of Neural Systems 17:231–240.

    https://doi.org/10.1142/S0129065707001093

    • Google Scholar
21. 1. Hafting T
    2. Fyhn M
    3. Molden S
    4. Moser MB
    5. Moser EI

    (2005) Microstructure of a spatial map in the entorhinal cortex

    Nature 436:801–806.

    https://doi.org/10.1038/nature03721

    • Google Scholar
22. 1. Histed MH
    2. Maunsell JH

    (2014) Cortical neural populations can guide behavior by integrating inputs linearly, independent of synchrony

    Proceedings of the National Academy of Sciences of USA 111:E178–E187.

    https://doi.org/10.1073/pnas.1318750111

    • Google Scholar
23. 1. Hunsberger E
    2. Scott M
    3. Eliasmith C

    (2014) The competing benefits of noise and heterogeneity in neural coding

    Neural Computation 26:1600–1623.

    https://doi.org/10.1162/NECO_a_00621

    • Google Scholar
24. 1. Jones R
    2. Buhl EH

    (1993) Basket-like interneurones in layer II of the entorhinal cortex exhibit a powerful NMDA-mediated synaptic excitation

    Neuroscience Letters 149:35–39.

    https://doi.org/10.1016/0304-3940(93)90341-H

    • Google Scholar
25. 1. Lerche H
    2. Jurkat Rott K
    3. Lehmann Horn F

    (2001) Ion channels and epilepsy

    American Journal of Medical Genetics 106:146–159.

    https://doi.org/10.1002/ajmg.1582

    • Google Scholar
26. 1. Lewis DA
    2. Curley AA
    3. Glausier JR
    4. Volk DW

    (2012) Cortical parvalbumin interneurons and cognitive dysfunction in schizophrenia

    Trends in Neurosciences 35:57–67.

    https://doi.org/10.1016/j.tins.2011.10.004

    • Google Scholar
27. 1. Longtin A
    2. Bulsara A
    3. Moss F

    (1991) Time-interval sequences in bistable systems and the noise-induced transmission of information by sensory neurons

    Physical Review Letters 67:656–659.

    https://doi.org/10.1103/PhysRevLett.67.656

    • Google Scholar
28. 1. Lundqvist M
    2. Compte A
    3. Lansner A

    (2010) Bistable, irregular firing and population oscillations in a modular attractor memory network

    PLOS Computational Biology 6:e1000803.

    https://doi.org/10.1371/journal.pcbi.1000803

    • Google Scholar
29. 1. Marder E

    (2012) Neuromodulation of neuronal circuits: back to the future

    Neuron 76:1–11.

    https://doi.org/10.1016/j.neuron.2012.09.010

    • Google Scholar
30. 1. Marder E
    2. Taylor AL

    (2011) Multiple models to capture the variability in biological neurons and networks

    Nature Neuroscience 14:133–138.

    https://doi.org/10.1038/nn.2735

    • Google Scholar
31. 1. Moser EI
    2. Moser MB

    (2013) Grid cells and neural coding in high-end cortices

    Neuron 80:765–774.

    https://doi.org/10.1016/j.neuron.2013.09.043

    • Google Scholar
32. 1. Onslow AC
    2. Jones MW
    3. Bogacz R

    (2014) A canonical circuit for generating phase-amplitude coupling

    PLOS ONE 9:e102591.

    https://doi.org/10.1371/journal.pone.0102591

    • Google Scholar
33. 1. Paoletti P
    2. Bellone C
    3. Zhou Q

    (2013) NMDA receptor subunit diversity: impact on receptor properties, synaptic plasticity and disease

    Nature Reviews. Neuroscience 14:383–400.

    https://doi.org/10.1038/nrn3504

    • Google Scholar
34. 1. Pastoll H
    2. Solanka L
    3. van Rossum MCW
    4. Nolan MF

    (2013) Feedback inhibition enables theta-nested gamma oscillations and grid firing fields

    Neuron 77:141–154.

    https://doi.org/10.1016/j.neuron.2012.11.032

    • Google Scholar
35. 1. Rubenstein JL
    2. Merzenich MM

    (2003) Model of autism: increased ratio of excitation/inhibition in key neural systems

    Genes, Brain and Behavior 2:255–267.

    https://doi.org/10.1034/j.1601-183X.2003.00037.x

    • Google Scholar
36. 1. Ruzzoli M
    2. Marzi CA
    3. Miniussi C

    (2010) The neural mechanisms of the effects of transcranial magnetic stimulation on perception

    Journal of Neurophysiology 103:2982–2989.

    https://doi.org/10.1152/jn.01096.2009

    • Google Scholar
37. 1. Sargolini F
    2. Fyhn M
    3. Hafting T
    4. McNaughton BL
    5. Witter MP
    6. Moser MB
    7. Moser EI

    (2006) Conjunctive representation of position, direction, and velocity in entorhinal cortex

    Science 312:758–762.

    https://doi.org/10.1126/science.1125572

    • Google Scholar
38. 1. Schmidt-Hieber C
    2. Häusser M

    (2013) Cellular mechanisms of spatial navigation in the medial entorhinal cortex

    Nature Neuroscience 16:325–331.

    https://doi.org/10.1038/nn.3340

    • Google Scholar
39. 1. Seung HS
    2. Lee DD
    3. Reis BY
    4. Tank DW

    (2000) Stability of the memory of eye position in a recurrent network of conductance-based model neurons

    Neuron 26:259–271.

    https://doi.org/10.1016/S0896-6273(00)81155-1

    • Google Scholar
40. 1. Shadlen MN
    2. Newsome WT

    (1994) Noise, neural codes and cortical organization

    Current Opinion in Neurobiology 4:569–579.

    https://doi.org/10.1016/0959-4388(94)90059-0

    • Google Scholar
41. 1. Shu Y
    2. Hasenstaub A
    3. Badoual M
    4. Bal T
    5. McCormick DA

    (2003)

    Barrages of synaptic activity control the gain and sensitivity of cortical neurons

    The Journal of Neuroscience 23:10388–10401.

    • Google Scholar
42. 1. Sigurdsson T
    2. Stark KL
    3. Karayiorgou M
    4. Gogos JA
    5. Gordon JA

    (2010) Impaired hippocampal-prefrontal synchrony in a genetic mouse model of schizophrenia

    Nature 464:763–767.

    https://doi.org/10.1038/nature08855

    • Google Scholar
43. 1. Skaggs WE
    2. McNaughton BL
    3. Wilson MA
    4. Barnes CA

    (1996) Theta phase precession in hippocampal neuronal populations and the compression of temporal sequences

    Hippocampus 6:149–172.

    https://doi.org/10.1002/(SICI)1098-1063(1996)6:2&amp;lt;149::AID-HIPO6&amp;gt;3.0.CO;2-K

    • Google Scholar
44. 1. Spellman TJ
    2. Gordon JA

    (2014) Synchrony in schizophrenia: a window into circuit-level pathophysiology

    Current Opinion in Neurobiology 30:17–23.

    https://doi.org/10.1016/j.conb.2014.08.009

    • Google Scholar
45. 1. Stensola H
    2. Stensola T
    3. Solstad T
    4. Frøland K
    5. Moser MB
    6. Moser EI

    (2012) The entorhinal grid map is discretized

    Nature 492:72–78.

    https://doi.org/10.1038/nature11649

    • Google Scholar
46. 1. Suthana N
    2. Haneef Z
    3. Stern J
    4. Mukamel R
    5. Behnke E
    6. Knowlton B
    7. Fried I

    (2012) Memory enhancement and deep-brain stimulation of the entorhinal area

    The New England Journal of Medicine 366:502–510.

    https://doi.org/10.1056/NEJMoa1107212

    • Google Scholar
47. 1. Tiesinga P
    2. Jose JV

    (2000) Robust gamma oscillations in networks of inhibitory hippocampal interneurons

    Network 11:1–23.

    https://doi.org/10.1088/0954-898X/11/1/301

    • Google Scholar
48. 1. Treiman DM

    (2001) GABAergic mechanisms in epilepsy

    Epilepsia 42:8–12.

    https://doi.org/10.1046/j.1528-1157.2001.042suppl.3008.x

    • Google Scholar
49. 1. Uhlhaas PJ
    2. Singer W

    (2012) Neuronal dynamics and neuropsychiatric disorders: toward a translational paradigm for dysfunctional large-scale networks

    Neuron 75:963–980.

    https://doi.org/10.1016/j.neuron.2012.09.004

    • Google Scholar
50. 1. Wang XJ

    (1999)

    Synaptic basis of cortical persistent activity: the importance of NMDA receptors to working memory

    Journal of Neuroscience 19:9587–9603.

    • Google Scholar
51. 1. Wang XJ
    2. Krystal JH

    (2014) Computational psychiatry

    Neuron 84:638–654.

    https://doi.org/10.1016/j.neuron.2014.10.018

    • Google Scholar
52. 1. Widloski J
    2. Fiete IR

    (2014) A model of grid cell development through spatial exploration and spike time-dependent plasticity

    Neuron 83:481–495.

    https://doi.org/10.1016/j.neuron.2014.06.018

    • Google Scholar
53. 1. Yoon K
    2. Buice MA
    3. Barry C
    4. Hayman R
    5. Burgess N
    6. Fiete IR

    (2013) Specific evidence of low-dimensional continuous attractor dynamics in grid cells

    Nature Neuroscience 16:1077–1084.

    https://doi.org/10.1038/nn.3450

    • Google Scholar
54. 1. Zhang K

    (1996)

    Representation of spatial orientation by the intrinsic dynamics of the head-direction cell ensemble: a theory

    The Journal of Neuroscience 16:2112–2126.

    • Google Scholar

Author details

1. Lukas Solanka

   1. Centre for Integrative Physiology, University of Edinburgh, Edinburgh, United Kingdom
   2. Institute for Adaptive and Neural Computation, Edinburgh, United Kingdom
   3. Neuroinformatics Doctoral Training Centre, School of Informatics, University of Edinburgh, Edinburgh, United Kingdom

   Contribution

   LS, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

2. Mark CW van Rossum

   Institute for Adaptive and Neural Computation, Edinburgh, United Kingdom

   Contribution

   MCWR, Conception and design, Analysis and interpretation of data

   Competing interests

   The authors declare that no competing interests exist.

3. Matthew F Nolan

   Centre for Integrative Physiology, University of Edinburgh, Edinburgh, United Kingdom

   Contribution

   MFN, Conception and design, Analysis and interpretation of data, Drafting or revising the article

   For correspondence

   mattnolan@ed.ac.uk

   Competing interests

   The authors declare that no competing interests exist.

   "This ORCID iD identifies the author of this article:" 0000-0003-1062-6501

• 1,940

  views

• 490

  downloads

• 28

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.