A Cortical Attractor Network with Martinotti Cells Driven by Facilitating Synapses

The population of pyramidal cells significantly outnumbers the inhibitory interneurons in the neocortex, while at the same time the diversity of interneuron types is much more pronounced. One acknowledged key role of inhibition is to control the rate and patterning of pyramidal cell firing via negative feedback, but most likely the diversity of inhibitory pathways is matched by a corresponding diversity of functional roles. An important distinguishing feature of cortical interneurons is the variability of the short-term plasticity properties of synapses received from pyramidal cells. The Martinotti cell type has recently come under scrutiny due to the distinctly facilitating nature of the synapses they receive from pyramidal cells. This distinguishes these neurons from basket cells and other inhibitory interneurons typically targeted by depressing synapses. A key aspect of the work reported here has been to pinpoint the role of this variability. We first set out to reproduce quantitatively based on in vitro data the di-synaptic inhibitory microcircuit connecting two pyramidal cells via one or a few Martinotti cells. In a second step, we embedded this microcircuit in a previously developed attractor memory network model of neocortical layers 2/3. This model network demonstrated that basket cells with their characteristic depressing synapses are the first to discharge when the network enters an attractor state and that Martinotti cells respond with a delay, thereby shifting the excitation-inhibition balance and acting to terminate the attractor state. A parameter sensitivity analysis suggested that Martinotti cells might, in fact, play a dominant role in setting the attractor dwell time and thus cortical speed of processing, with cellular adaptation and synaptic depression having a less prominent role than previously thought.


Introduction
The population of inhibitory interneurons comprises only 15-20% of neocortical neurons. Despite being a minority, interneurons are believed to play an important role in shaping network activity patterns by directly controlling the input/output of principal cells. Recent advances in single-cell recording techniques and molecular biology have led to an explosion of data and knowledge about these inhibitory cells -their morphology, synaptic connections, short-term plasticity and molecular characteristics [1][2][3]. Much less is known about the specific functional roles in network function played by the diverse inhibitory interneuron subtypes. Apart from the differences in morphology and synaptic targeting on pyramidal cells (PC), interneuron synapses also exhibit different synaptic short-term plasticity properties, from strongly depressing to strongly facilitating. Synaptic depression of glutamatergic synapses between pyramidal cells has a dominant effect in controlling firing rate [4]. Only very few cortical models of various phenomena like gamma oscillations, working memory, slow-wave sleep (SWS) oscillations etc take into account this dynamical nature of the synapses [5]. Most of them add details to the cell morphology keeping synapses, by contrast, static [6][7][8][9][10].
We have previously developed and characterized a network model of neocortical layers 2/3. This cortical network model has a modular hypercolumnar structure in which each hypercolumn comprises a set of minicolumns. Such a module operates like a soft winner-take-all network typically allowing just one active minicolumn at a time. It operates as an attractor type associative memory and displays bistable irregular low-frequency firing and various phenomena like, e.g. pattern retrieval, completion and rivalry as well as spontaneous wandering of network between stored states [11][12][13]. Mounting experimental evidence shows that ongoing activity in cortex can exhibit complex spatiotemporal patters [14][15][16]. These patterns seem to wander among a set of intrinsic cortical states that reflects the overall cortical architecture. Also, using voltage sensitive dye imaging Kenet et al. revealed that in primary visual cortex these cortical states matched the functional map of orientation columns [17]. We here demonstrate how adding one type of inhibitory interneuron, the Martinotti cell (MC), to our previous network model affects the attractor network dynamics during spontaneous reactivation. A computational study undertaken by Melamed et al. (2008) [18] showed that slow oscillations emerge from the interplay of excitatory and inhibitory populations. Like MCs, their inhibitory population received facilitating synapses from the excitatory population and it was demonstrated that such inhibition underlies the switching between up and down states. This study was based on a rate-based model with non-adapting excitatory cells connected by static synapses and utilized only one type of interneuron. A recent experimental study on cortical slices has also suggested that interneurons receiving facilitating synapses in neocortex might play a key role in the termination of up states [19].
The objective of the present study is to study the impact of inclusion of a late-onset interneuron, that is MC, in our attractor network model by investigating how it affects activity levels and attractor dwell time during spontaneous activity. Although an earlier study has shown the effects of late-firing MC in a firing rate model, ours is the first study that uses spiking units with PC-MC characteristics matched to in vitro data. Here we use single-compartmental Hodgkin-Huxley (HH) type neuron models [20] of two different types of interneurons embedded in a population of pyramidal cells with the ratio of pyramidal to interneuron being 90:10. We have included dynamic synapses throughout the network, enabling us to show how pyramidal cells differentially excite interneurons via depressing and facilitating synapses. We commenced with reproducing in our model the PC-MC sub-circuit, as previously described by Silberberg & Markram (2007) and Silberberg (2008) [21,22], and reproduced (a) frequency dependent disynaptic inhibition of pyramid cells, and (b) frequency dependent recruitment of MCs. Thereafter, we integrated this microcircuit in our cortical attractor network model [11][12][13] to study the dynamic effects on a more global scale. We addressed the effect of MCs on the attractor dwell time when the network operated without external input, thus freely ''hopping'' between the stored states. We show that basket cells (BC) that receive depressing excitatory synapses have a high firing rate at the beginning of the attractor state which then tapers off. On the other hand, MCs that receive facilitating synapses display a late onset of activation and tend to terminate an ongoing attractor state. We have further shown how the dwell time and peak firing of PCs varies with PC-MC connection density. An earlier computational study [12], in agreement with others [9,23,24,25,26,27] had demonstrated spike-frequency adaptation and synaptic depression contributing to termination of attractor states. We here show how the MC activation could have an even stronger contribution to the termination of the attractor states relative to spike-frequency adaptation and synaptic depression between PCs.

Model Neurons
The cells included are layer 2/3 pyramidal cells (PC) and two different types of inhibitory interneurons. They are soma targeting horizontally projecting basket cells (BC) [28] and dendrite targeting, vertically projecting Martinotti cells (MC) that establish a disynaptic inhibitory feedback pathway between the pyramidal cells [22,29].
PCs are of regular firing type. In the previous simulations, adaptation was modeled by calcium entering via voltage gated Cachannels and activation of K Ca channels. Here adaptation is modeled using the M-current, a slow non-inactivating potassium current described by Yamada et al. [30]. BCs are modeled as nonadapting, relatively fast-spiking cells. MCs have the same properties as BCs except that they are somewhat more adapting [22]. The multi-compartmental cells used in the previous work have here been replaced by single-compartment cells with size of each cell type's soma, steady-state current and voltage equations, and conductance values taken from Pospischill et al. 2008 [31] ( Table 1).
All models described here were single-compartment neurons (cylinder of diameter d and length L) described by the following membrane equation: membrane potential, C m = specific capacitance of the membrane, g leak = specific resting (leak) membrane conductance, E leak = resting membrane reversal potential. The kinetic parameters of the voltage-dependent Na current is given by where g Na and E Na of different cortical cells are given in Table 1.
The kinetic parameters of the voltage-dependent K (delayed rectifier) current is given by where g Kd and E Kd of different cortical cells are given in Table 1.
The kinetic parameters of the voltage-dependent M current is given by where g M and t max of different cortical cells are given in Table 1.
We have used a point-conductance model of synaptic noise to account for the stochastic variation of conductance due to synaptic background activity on all cell models [32]. Table 2 gives parameters for mean conductance (g e0 and g i0 ) and standard deviation (s e and s i ). The level of this background noise is adjusted to a low firing rate on all cells (0.25 -0.5 Hz).

Model Synapses
Glutamatergic synapses work on two broad categories of receptors: kainate/AMPA and NMDA. A mix of both provides the PC-PC glutamatergic transmission, but the PC-BC glutamatergic transmission is purely kainate/AMPA [33]. It is inconclusive from experiments whether PC-MC glutamatergic tranmission is plainly kainate/AMPA or a mix. For simulations presented in this paper, it is entirely kainate/AMPA. The GABA-ergic transmission in our model is exerted solely by GABA A [33] (See Table 3). AMPA and GABA A currents are given by [34,35]: where the gating variable s (the fraction of open channels) is described by first-order kinetics via two equations: The NMDA current is given by: The gating variable s obeys the same types of equations (1,2). We have taken t x = 0.05 ms and t s = 6 ms for AMPA and GABA A , t x = 5 ms and t s = 150 ms for NMDA, a x = 1 (dimensionless) and a s = 1 (ms 21 ) for AMPA, NMDA and GABA A [5,35]. Short-term depression and facilitation were incorporated for all glutamatergic and GABAergic synapses [36,37]. Every presynaptic spike, occurring at time t sp , causes a fraction U of the available pool to be utilized, the rate of return of resources given by t rec , is multiplied by a quantity R (the fraction of available vesicles). R obeys the dynamical equation [38]: The short-term depression is introduced into the synapse model by multiplying a x in (1), which mimics the transmitter release per spike, by R in (3) which is the fraction of available vesicles.
In modeling a facilitating synapse, U becomes a dynamic variable increasing at each presynaptic spike and decaying to the baseline level in the absence of spikes.
where U1 is a constant that determines the step increase in U and t facil is the decay time constant of facilitation. At most three parameters completely define each connection type; U, t rec and t facil (depressing) or U1, t rec and t facil (facilitating). On one hand, there is no consensus on how precise these values should be. On the other hand, experiments do conform on a range of values [1]. The traces and parameters fitted to the model provided by Gilad Silberberg from his own experimental studies were very useful in setting these values. Particularly useful were the short-term dynamics between PC -MC and MC -PC connections. A connection with a high 'U' factor means the synapse has a strong 'postsynaptic punch' for initial spikes followed by a rapid depression. A low 'U' factor means lower initial release probability and thereby low initial impact on the postsynaptic side saving the transmitters for future presynaptic spikes. A synapse's effect can be strongly depressing (t rec .. t facil ) or strongly facilitating (t rec ,, t facil ) or intermediate displaying combined depressing-facilitating behaviour. The values assigned for each connection can be seen in Table 4.
We used Thomson

Simulation
The computation model was run using the NEURON simulator [40]. Simulations were typically performed on 128 nodes of the Blue Gene/L computer at the Center for Parallel Computers at KTH. It took 70 seconds to simulate one second of network activity.  [12,13,41]. The sub-sampled neocortical model used here represents a 363 mm patch of cortex arranged on a square topology of 666 hypercolumns each separated by 500 mm, in agreement with hypercolumn diameter data from cat, i.e., 300-600 mm [42]. Each hypercolumn further constitutes several minicolumns -various estimates suggest that there are about one hundred minicolumns bundled into a hypercolumn [43]. In the current sub-sampled network model we have 5 minicolumns. The arrangement of cells in the local microcircuit together with connection probabilities and strengths (PSP amplitudes) are shown in Fig. 2. Each minicolumn (red disc) consists of 30 PCs densely connected to other PCs in the same minicolumn (25%) [44,45,46] and two regular spiking non-pyramidal (RSNP) interneurons (possibly double-bouquet cells) (not shown). Each hypercolumn has 8 BCs (circular blue disc). Each PC in a minicolumn targets 70% of its neighboring BCs and each BC targets 70% of the neighboring PCs. In the recent versions of the model, we have also introduced synaptic connections between BCs (40%) in the same hypercolumn. Even though electrical coupling between BCs have been observed, we have not included those in our model [47,48]. The long-range minicolumn -minicolumn inhibition through RSNP cells, used in the previous study [12], is turned off here, since the dwell time of the attractor, which we measure here is not affected by its presence. Our new addition to this model is the MC pool (3 per hypercolumn, oval blue disc). The PC -. MC and MC -. PC connections show high convergence and divergence [49]. In our model, each MC receives input from 40% of the PCs in the hypercolumn and contacts 80% of the PCs in the hypercolumn [22,49]. The extent of BC and MC inhibition is limited to the home hypercolumn. The minicolumns in a hypercolumn altogether sweep a width of approximately 100 mm and hence we have made two assumptions. (a) The extent of inhibition of horizontally projecting BCs and vertically projecting MCs may vary in the real cortex, but in this subsampled network, all the minicolumns lie within the reach of the BC and MC pool. (b) The PCs in each minicolumn target 8 neighboring BCs and 3 neighboring MCs. Since each hypercolumn has closely spaced adjacent minicolumns, all minicolumns share the same BC and MC pool. If we had not clustered minicolumns like it is done here, we could not have assumed just one pool.

Architecture of the Network Model
The cartoon in Fig. 2 shows how the minicolumns in different hypercolumns, denoted by dashed lines, are connected. Thus, a set of minicolumns distributed over different hypercolumns represents a stored pattern or memory or an attractor of the network dynamics. In each hypercolumn, via lateral inhibition of BCs, the activity in an attractor state engages only one minicolumn (orthogonal patterns). In this network, consequently, we store as many patterns as the number of minicolumns in a hypercolumn. But by allowing overlapping memory patterns the number of patterns stored can be increased significantly [50].

PC -MC Microcircuitry
We set out to reproduce how discharge of an individual PC at different rates induces differential delays in the discharge of MCs and how this influences a second PC. Fig. 1a shows the connection setup of this disynaptic inhibitory pathway involving two neighboring PCs and three intermediate MCs. High frequency activation of PCs is shown to exert inhibition in a significantly larger number of PCs by a supralinear increase in the recruitment of MCs [29]. To this end, we included synaptic background activity to show frequency dependent recruitment of MCs.
Connections from PCs to MCs are facilitating (U1 = 0.05, t rec = 20 ms, t facil = 1000 ms) [18]. Presenting an AP train in a presynaptic PC1 thus evokes a discharge in a post-synaptic MC with a delay and a delayed inhibition of PC2, which is postsynaptic to the MC. The latency of this discharge onset depends upon presynaptic stimulation frequency, as shown in slices [22]. Since the PC -MC connections are facilitating, they are endowed with low initial release probability, long facilitation time constants, short depression time constants (see Table 4 Connections from MCs to PCs displayed synaptic depression (U = 0.3, t rec = 500 ms, t facil = 5 ms) [1]. Monosynaptic IPSPs of these connections, as reported, have a small amplitude (-0.6 mV). Disynaptic responses on PC2 for different frequencies (40, 50 and 70 Hz) are shown in Fig. 1c. It demonstrates how the disynaptic responses of the model neuron increased in amplitude and decreased in latency as a function of presynaptic AP train frequency in accordance with Silberberg & Markram (2007) [22]. The change in amplitude of the disynaptic response from a 40 Hz presynaptic AP train to a 70 Hz AP train was reported nearly twice as high and intermediate for a 50 Hz AP train. In our model, we were able to fit this relationship quantitatively, in the presence of a jitter, when we assumed three MCs.

MCs in a Network
We were then interested to investigate the effect of including MCs in our earlier attractor network model [12]. Previous studies showed how this network is capable of performing basic attractor network operations like pattern completion and pattern rivalry. Our focus now is to investigate the dynamic effects of MCs on the network activity, in the absence of any sensory-like external input.
Initially, we performed simulations omitting MCs. When the network was subjected to a low level of background noise, it started visiting various stored states randomly (Fig. 3a, top). The raster plot clearly shows how the PCs in different hypercolumns that form a pattern were (nearly) synchronously active. The rate of switching is about three states per second. The neurons that are active when the network engages in an attractor state receive stronger synaptic input, raised average membrane potential and an increase in spike rate, in agreement with the previous studies [12]. These concurrently active cells through local recurrent connections and global long-range connections maintain the persistent activity. Mean firing rate is a good measure to ascertain if the network has indeed entered an attractor state, and inquire into time evolution of PC and BC firing rates (Fig. 3a, bottom). It is calculated by setting a time window of 25 ms and counting the number of spikes that occurred in this time window, dividing this by the length of the time window and the number of active PCs. PCs fire (blue curve) briskly at the beginning of the attractor visit and the firing rate decreases owing to spike-frequency adaptation and synaptic depression (see below). Termination of an active attractor gives way to the activation of silent attractors, this cycle repeats ad infinitum. BCs fire at every attractor cycle since they are connected to all the minicolumns. Also, the BCs adjust their firing (red curve) in synchrony with PCs showing how excitation and inhibition during cortical activity is always in a state of dynamic balance [51].
MCs were introduced using the following connection paradigm: Each MC receiving input from 40% of PCs in a minicolumn and projecting back to 80% of PCs in a minicolumn. Below we discuss different connection paradigms. With the inclusion of MCs, the attractor state experiences early termination (Fig. 3b, top) resulting in an increased frequency in state transitions (from 2 Hz to 5 Hz). This early termination is due to delayed inhibition provided by MCs. The raster plot is similar to above, with spikes from MCs (green dots) included. The late firing of MCs is not conspicuous in the raster plot, so we instead calculated average discharge rates (Fig. 3b, bottom). The BCs that receive depressing synapses and high unitary EPSPs are the first to respond and provide inhibition throughout the attractor state (red curve) just like above. The MCs that receive low unitary EPSPs, characteristic of facilitating synapses, starts to engage with a delay after the onset of attractor activity (green curve), and thereby sweeping away the activity [52]. This is due to high divergence of MCs on its neighboring PCs, also shown in slices [22,49]. Similar to BCs, MCs also fire at every attractor cycle since they are shared by all minicolumns in a hypercolumn.

Different PC to MC Connection Strategies
MCs can be integrated into our network model in different ways compatible with experimental data. Here we investigated three different connection paradigms between PCs and MCs. We kept the MC R PC connections, which shows high divergence (80%), constant varying only the PC R MC connections. The different connection paradigms are (Con/Div): (15/80, 40/80, 80/80), where Con stands for convergence, and Div stands for divergence. For instance, if we assume the 40/80 paradigm, each MC receives converging input from 40% of presynaptic PCs in a minicolumn and diverges onto 80% of postsynaptic PCs in a minicolumn. If a PC exerts inhibition on another PC via 1 MC, 2 MCs and 3 MCs, we call it Type 1, Type 2 and Type 3 respectively. Going back to the microcircuit in Fig. 1, we find Type 1, Type 2 and Type 3 when the input frequency is 40, 50 and 70 Hz respectively. Table 5 contains the average number of Type 1, Type 2 and Type 3 connections in each connection paradigm.
Increasing the PC -PC synaptic strength decreased the attractor duration, owing to increase in firing and faster adaptation and depletion of vesicles (Fig. 4a). The response in that ''15/80'' paradigm was similar to the network without MCs since 15% convergence of PCs on MCs did not lead to much MC discharge. However, ''40/80'' and ''80/80'' paradigms had distinguishable effect due to higher convergence (Fig. 4a). Increasing the PC -MC connection strength showed a similar effect (Fig. 4c), for obvious reasons. When PCs enter the attractor state, their firing rate will be on the increase for about 100 ms, until nonlinearities like synaptic depression and spike-frequency adaptation kicks in bringing the firing rate down. If the discharge of MCs supersedes the above nonlinearities, the peak firing rate will be lowered, otherwise it remains unaffected as shown in Fig. 4c, d. The values that were held constant are shown on the top of each sub-figure.

Regulation of Attractor Dwell Time
The attractor state termination could be due to intrinsic neuron properties like adaptation and synaptic depression or due to the presence of MCs inhibition in the network. At this juncture, we were interested to investigate which factors dominate in controlling the attractor state dwell time. To address the effect of synaptic depression, we varied the strength of depression between PC -PC synapses by lowering t rec that would speed up the rate of recovery of vesicles and by lowering U that would decrease the vesicle release probability resulting in lower depression [36]. We controlled the effect of spike-frequency adaptation by down-  regulating the conductance of the M-current (g M ). Figure 5 shows the attractor duration calculated for various conditions in the presence and absence of MC inhibition. In the absence of MC inhibition (blue curve), at lower values of t rec and U, PCs discharge more due to lesser depression, leading to an increase in attractor duration in comparison with the control condition. At higher values of t rec and U, the attractor duration attains constancy. In the same vein, lowering g M of PCs prolongs the attractor state conforming to previous studies [12,41].
In the presence of MC inhibition (Fig. 5, red curve), the attractor duration is reduced as shown in the control condition 'C'. However, the lower values of t rec and U causing lesser depression did not prolong the attractor state as it did when the MC inhibition was turned off. Furthermore, the variation in the dwell time at low t rec and U, shown by the height of the error bar, in the presence of MCs is low in comparison to when MCs inhibition was absent. As expected, lowering g M of PCs in the presence of MCs did not change the attractor duration significantly. Comparing the red curve and green curve in Figure 5 demonstrates that the network shows similar response for a change in t facil . But large changes in PC-MC or MC-PC connection density affected the attractor duration (not shown here). Thus, in this network model the MC inhibition is the dominating factor in regulating the attractor dwell time and it brings more stability at lower values of depression and spike-frequency adaptation. The latter aspect becomes important in the light of heterogeneity in short-term synaptic parameters.  Fig. 6). Having BCs alone in the hypercolumn (blue and green) makes the dwell time amenable to changes in short-term synaptic and cell intrinsic parameters. It is only due to the presence of MCs (red) that the dwell time is insensitive to these changes.

Discussion
In this work, we commenced by modeling the PC-MC microcircuit, as previously shown by Silberberg et al. [22], and reproduced (a) frequency dependent disynaptic inhibition of PCs (b) frequency dependent recruitment of MCs. The model microcircuit contained 3 MCs mediating disynaptic inhibition between 2 PCs. The PC -MC synapses were facilitating and MC -PC synapses were depressing. We stimulated the presynaptic PC with trains of AP at different frequencies and demonstrated increase in amplitude and decrease in latency of disynaptic response of the post-synaptic PC (Fig. 1). Real neurons in the brain, excitatory and inhibitory alike, are constantly bombarded with synaptic inputs causing the membrane potential to fluctuate. In the absence of any jitter, the number of MCs responding to firing of presynaptic PC at different frequencies would remain constant. When a Poisson source was added to all MCs, we observed trial-to-trial variability in the response of MCs. Only one MC fired in most of the trials when the input frequency was 40 Hz, about 2 MCs fired in 90 percent of trials for 50 Hz and all MCs discharged reliably when the input frequency was 70 Hz (Fig. 1). Our model MCs displayed frequency dependent recruitment much similar to slice data and we managed to fit disynaptic response quantitatively using 3 MCs [22,29]. We then integrated this PC-MC microcircuit in our cortical attractor network model (Fig. 2) [11][12][13]. Amidst the array of phenomena that our network model can exhibit, we focused on the effect of MCs when the network is spontaneously hopping between the attractor (memory) states in the absence of any external input. Such attractor states have been observed in neocortical slices with their active state duration ranging typically from 50 ms to 5 s [16]. In our model, as described below, this dwell time can be controlled by various parameters. Raster plots and average firing rates were used to demonstrate the temporal variation in the firing of BCs and MCs. We showed that BCs that receive depressing synapses has a high firing rate at the beginning of the attractor state which then tapers off. On the other hand, MCs that receive facilitating synapses display a late onset of activation and promote termination of the attractor states owing to their high divergence onto the PCs (Fig. 3). In our model the excitation-inhibition dynamic balance during the attractor state is maintained by the PC -BC pathway until the discharge of MCs tips this balance by elevating inhibition.
Our earlier computational study [12], in agreement with others [9,53], had demonstrated spike-frequency adaptation and synaptic depression contributing to termination of attractor states. A computational model in the context of relaxation oscillations has recently addressed whether cellular adaptation or synaptic depression affects episode inititation/termination the most [27]. Our attractor dwell time also relies upon those factors in the absence of MCs, but the presence of MCs makes the dwell time impervious to those changes (Fig. 5). Our network model overtly demonstrates that the late onset of MC discharge due to facilitating synapses has a stronger contribution in setting the attractor dwell time and rate of switching between the attractor states compared to spike-frequency adaptation and synaptic depression between PCs.
Introducing three synaptically different interneurons (MC) in our hypercolumn increased the robustness of our network as described above. In order to assay if it is the presence of MCs that lead to increase in robustness, we replaced the three MCs with three BCs in each hypercolumn and performed the sensitivity analysis varying the same set of parameters (Fig. 6). Clearly, the ''w/o MC'' and ''BC(11)'' exhibited the same behaviour for changes in synaptic and adaptation parameters. Thus the presence of temporally different interneurons, early-onset BC and late-onset MC, did make the network more robust.
We find gamma oscillations when the network is in the attractor state. We haven't shown the synthetic LFP spectrograms here but this result could be extended from the analysis made in the previous work [13]. These gamma oscillations in every attractor state are caused by the PC-BC reciprocal circuitry in conformity with the well-established idea that entrainment of PCs by BCs causes gamma oscillations [54,55]. The dwell time of our attractor states with MC inhibition is about 200 ms, which results in a (5 -6) Hz state transition frequency. We propose that it is the combined action of BC and MC inhibition that causes this theta/ gamma oscillation. With BC inhibition alone attractor switching remains but gets slower (Fig. 5, 6). When we replace all the BCs with MCs, the attractor dynamics were largely disrupted (data not shown). There was a period of no inhibition on all the PCs due to the late onset of MC discharge during which all the patterns were active and this also affected the theta/gamma phenomenon.
When quiescent neocortical slices are bathed in artificial cerebrospinal fluid (ACSF), it induces spontaneous fluctuations between two quasi-stable states knows as UP-and DOWN-states [53]. The stability and repeatability in the patterns of firing during UP states have suggested that they could correspond to attractor states of the network [16]. By using new optogenetic techniques [56][57][58] it should be possible to selectively depolarize or hyperpolarize the different types of inhibitory cells that participate in this activity and investigate the roles of each type. Our model proposes that light-activation of MCs would terminate UP-states while hyperpolarization would extend them. Selective lightinduced depression of BCs would likely lead to more vigorous and less distinct oscillatory network activity.
As mentioned in the introduction, a model proposed by Melamed et al. [18] attributes the interaction of excitatory neurons with recurrent excitation and interneurons receiving facilitating synapses as the basis for the generation of slow oscillations in the neocortex. They showed the dependence of network oscillations on the strength of PC-MC synapses corroborating with our study (Fig. 4). However, our study showed no change in the attractor rate switching for any change in the facilitation time constant (Fig. 5), which is not in agreement with their study. The main differences between the two models are the type of neuron model and the division of labour between interneurons. Our spiking model based on the Hodgkin-Huxley formalism utilized two interneurons with temporally disparate workings, whereas their rate-based model only represented the MC interneuron population. Despite these differences, both models pinpoint the decrease in excitation and elevation of inhibition as the most plausible causes of attractor termination. Some recent experimental studies had addressed the receptive field properties of excitatory and inhibitory cells in layer 2/3 and 4 of mouse visual cortex [59,60]. The orientation tuning and direction selectivity of pyramidal cells were highly tuned as opposed to basket cells that were largely unselective to any particular orientation concurring well with our modeling study. Interestingly, MCs were also direction selective yet not as highly tuned as PCs and provided delayed inhibition. In the small-scale version of our network presented here, the extent of BC and MC inhibition is the same since each hypercolumn has only 5 minicolumns. But in our full-scale model the number of minicolumns receiving BC inhibition will be more than those receiving MC inhibition since BC inhibition is more laterally spread than MC inhibition in the real cortex. Thus in our full-scale conceptual model, the orientation tuning of MCs would be intermediate between PCs that are highly selective and BCs that are largely unselective, concurring well with the experimental study.
The short-term plasticity parameters used in this model are taken from silent slice data [1,22], but the cortex is permanently active during awake and sleep states. Thus the presence of ongoing cortical activity might have an impact on short-term plasticity. Some experiments show this to be the case. Synaptic potentials showed a stronger depression in silent slices than in the active cortical network in vivo and in vitro [61,62]. From our sensitivity analysis, it is clear how attractor dwell times change in relation to short-term depression parameters. Future in vivo experiments describing synaptic pathways of excitatory and inhibitory neurons need to take this difference into consideration.