Analysis of an Attractor Neural Network’s Response to Conflicting External Inputs
Abstract
The theory of attractor neural networks has been influential in our understanding of the neural processes underlying spatial, declarative, and episodic memory. Many theoretical studies focus on the inherent properties of an attractor, such as its structure and capacity. Relatively little is known about how an attractor neural network responds to external inputs, which often carry conflicting information about a stimulus. In this paper we analyze the behavior of an attractor neural network driven by two conflicting external inputs. Our focus is on analyzing the emergent properties of the megamap model, a quasicontinuous attractor network in which place cells are flexibly recombined to represent a large spatial environment. In this model, the system shows a sharp transition from the winnertakeall mode, which is characteristic of standard continuous attractor neural networks, to a combinatorial mode in which the equilibrium activity pattern combines embedded attractor states in response to conflicting external inputs. We derive a numerical test for determining the operational mode of the system a priori. We then derive a linear transformation from the full megamap model with thousands of neurons to a reduced 2unit model that has similar qualitative behavior. Our analysis of the reduced model and explicit expressions relating the parameters of the reduced model to the megamap elucidate the conditions under which the combinatorial mode emerges and the dynamics in each mode given the relative strength of the attractor network and the relative strength of the two conflicting inputs. Although we focus on a particular attractor network model, we describe a set of conditions under which our analysis can be applied to more general attractor neural networks.
Abbreviations
 WTA
winnertakeall mode. The WTA mode refers to an operational mode of the dynamical system in which any stable fixed point corresponds to exactly one activity bump.
 Comb.
combinatorial. The comb. mode refers to an operational mode of the dynamical system in which there exist stable fixed points corresponding to multiple activity bumps.
1 Introduction
The theory of attractor neural networks has greatly influenced our understanding of the mechanisms underlying the computations performed by neural networks. This is especially true for hippocampal networks involved in spatial, declarative, and episodic memory. According to this theory, structured recurrent connections among N neurons cause the Ndimensional state vector to converge in time to a stable, lowdimensional space called the attractor [1]. Such a network embeds memories as stationary attractors, which may be a discrete set of point attractors representing a discrete set of objects [2] or a continuum of attractor states representing continuous variables such as heading direction [3, 4] or spatial location within an environment [5, 6, 7, 8, 9, 10]. Numerous theoretical studies have revealed properties of attractor neural networks that make them a desirable neural mechanism for memory storage, such as robustness to damage, pattern completion, and generalization [11, 12]. Attractor neural networks should arise naturally in regions of the brain with recurrently connected neurons and Hebbiantype synaptic plasticity, and they provide a theoretical framework for experimental design and data interpretation [13].
Attractor neural networks have been studied extensively through both analysis and computational simulations [1, 3, 14, 15, 16, 17]. While some studies do examine the role of external input [16, 18, 19], most determine the set of stable equilibrium states in the absence of external input, establishing properties such as the structure and capacity of the attractor. Relatively little is known about how an attractor network may respond to conflicting external inputs. This creates a gap between the idealistic predictions of attractor network theory and experimental data, since it is often experimentally difficult if not impossible to isolate putative attractor dynamics from the influence of the strong (often conflicting) external inputs into the neural network. In the current study, we analyze an attractor neural network’s response to conflicting external inputs that effectively create a competition between embedded attractor states. Our focus is the interesting behavior observed in our numerical simulations of the megamap model, a quasicontinuous attractor network representing a large spatial environment, driven by external inputs encoding two different locations in the environment [10]. However, the analytical methods and results obtained here can be applied to more general attractor network models.
The megamap model is designed for a network of principal cells in the CA3 subregion of the hippocampus, a region crucial for learning and memory [20, 21, 22]. These cells are often referred to as place cells due the strong spatial correlate of their activity. In small, standard recording environments (∼1 m^{2}), a given place cell is primarily active when the animal is within one specific subregion of the environment, called the cell’s place field [21, 23]. The megamap model flexibly recombines place cells to extend standard attractor network models of place cells, in which the majority of cells have one place field, to larger environments in which place cells have been shown experimentally to have multiple, irregularly spaced place fields [24, 25, 26]. The model follows logically from the recurrent connections among place cells in the CA3 [27], the Hebbianlike associative plasticity observed in the hippocampus [12, 28, 29], and the consistent coactivity of place cells with neighboring place fields [30].
Since the megamap seamlessly represents much larger environments than is possible for standard attractor network models of place cells, it allows us to explore whether any interesting dynamics emerge in large environments. In our numerical simulations, we observed a sharp transition in the network’s response to conflicting external inputs as the environment continuously grew in size [10]. In relatively small environments, the megamap behaves similarly to standard continuous attractor neural networks, operating in the winnertakeall (WTA) mode whereby the equilibrium state fully represents one input while effectively ignoring the second input. In larger environments, the megamap operates in the combinatorial mode, effectively combining two embedded attractor states to stably represent both inputs. Furthermore, we observed hysteresis, a classic characteristic of attractor dynamics, in the WTA mode, but the initial state had no effect on the equilibrium state in the combinatorial mode. The combinatorial mode is an interesting emergent property of the model that may be related to the partial remapping of hippocampal place cells sometimes observed when an animal is introduced to a new environment that simultaneously resembles two different familiar environments. In this cue conflict situation, the evoked neural responses are often mixtures of the responses to both environments rather than representations of one environment only [31]. The combinatorial mode emerges in the megamap model in sufficiently large environments when the weights are set optimally through gradient descent but not when the weights are set by the basic Hebbian learning rule [32, 33]. The latter method is widely used in attractor network models of place cells representing multiple environments [5, 6, 34, 35, 36].
We previously explored this emergent property of the megamap through numerical simulations and discussed its implications [10]. In the current study, we use mathematical analysis to derive a numerical test for determining the operational mode of the system a priori, characterize the conditions under which the combinatorial mode emerges, and derive explicit equations for the parameters of the model at which bifurcations occur. The numerical test is derived through stability analysis. It is an easily applied, useful tool for determining the expected response of a general attractor network to conflicting external inputs. This is particularly useful when the attractor network is selforganized. The latter two results are obtained through a linear mapping of the Ndimensional dynamical system to a 2dimension reduced model. Analysis of the stable fixed points of the reduced model elucidates the attractor network strength, which we quantify, and the relative strength of conflicting external inputs for which the equilibrium state vector represents the first location, represents the second location, represents one location or the other dependent on the initial state (hysteresis), or represents both locations. The explicit equations relating the dynamics of the attractor network to the model parameters are particularly useful when designing an attractor network to model a set of observed phenomena.
An outline of the paper is as follows. In Sect. 2, we present the dynamical system of the megamap model and describe two methods used to set the recurrent weights. We then show a numerical example of the operational modes and derive a numerical test for determining the operational mode. In Sect. 3, we present the reduced 2unit model and describe the conditions under which the reduced model is an accurate approximation of the full attractor network model. In Sect. 4, we characterize the conditions under which the combinatorial mode emerges and derive equations for the bifurcations of the dynamical system. We close in Sect. 5 by comparing our analysis to other analytical treatments of attractor neural networks, describing possible extensions of the reduced model, and discussing the implications of the results for various types of attractor network models.
2 Operational Modes of the Megamap
We begin by describing the basic equations governing the megamap model and by illustrating the operational modes through a numerical example. For further details, see [10].
2.1 Megamap Model
 (1)The optimal weights are set incrementally through the delta rule [33] so that a set of desired activity patterns, \(\{\overline {\textbf{f}}(\textbf{x}_{j})\}\), are embedded into the network as stable fixed points of the dynamical system (Eq. (1)) when the external input into each cell is an idealistic sum of Gaussians centered at the preferred locations of each cell (Fig. 1(a)). The desired activity of each cell is the sum of Gaussianlike place fields. Explicitly, for each cell i with \(M_{i}\) place fields centered at \(\{\textbf{c}_{im}\}_{m=1}^{M_{i}}\), the training input and desired activity are, respectively, given bywhen the animal is stationary at location x. The training input is set as the idealistic sum of Gaussian bumps whose amplitudes are given by the parameter \(\overline {b_{\mathrm{pk}}}\). The desired activity is set as the sum of activity bumps of height \(f_{\mathrm{pk}}\) over each place field center. The shift parameter, \(u_{0}>0\), is the depolarization at which a cell becomes active. The optimal weights are set using a discrete set of locations \(\{\textbf{x}_{j}\}\) distributed uniformly over the environment (at least 15 cm from a boundary).$$ \begin{aligned} \overline {b}_{i}(\textbf{x}) &= \overline {b_{\mathrm{pk}}} \sum_{m=1}^{M_{i}} \exp\biggl(\frac{\textbf{x}\textbf{c}_{im}^{2}}{2\sigma^{2}} \biggr)\quad\mathrm{and}\\ \overline {f}_{i}(\textbf{x}) &= \sum_{m=1}^{M_{i}} f \biggl( (1+u_{0})\exp\biggl(\frac{\textbf{x}\textbf {c}_{im}^{2}}{2\sigma^{2}} \biggr)  u_{0} \biggr) \end{aligned} $$(2)
 (2)The Hebbian weights are set as the sum of tuning curves,where each cell j has the preferred locations \(\{\textbf{c}_{jm}\} _{m=1}^{M_{j}}\), and \(w_{\mathrm{tune}}\) is the weight profile determined by computing the optimal weights when each cell has exactly one place field. This tuning curve is approximately Gaussian, and setting weights as the sum of Gaussians is a common method for constructing attractor network models of place cells [5, 6, 34, 35, 36]. The resulting weights approximate the weights expected given the basic Hebbian learning rule [32, 33, 34].$$W_{jk} = W_{kj} = \sum_{m=1}^{M_{j}} \sum_{n=1}^{M_{k}} w_{\mathrm{tune}}\bigl( \textbf{c}_{jm}\textbf{c}_{kn}\bigr), $$
2.2 Numerical Example of the Operational Modes of the Megamap
Since the megamap can seamlessly represent much larger environments than was previously possible, the model allows one to explore whether any interesting properties emerge when the attractor network represents a large space. We found that the megamap with optimal weights sharply transitions from a winnertakeall (WTA) mode to a combinatorial mode as the environment becomes sufficiently large [10]. While a megamap in either mode is similarly robust to a noisy or incomplete external input, there are clear differences between the modes when the network is driven by conflicting external input encoding multiple locations in the environment. In this situation, small megamaps operating in the WTA mode effectively suppress the input encoding one location and fully represent the second location, but large megamaps operating in the combinatorial mode robustly represent both locations through two costable activity bumps (Fig. 1(c) and (d)). Moreover, hysteresis is observed only in the WTA mode, and a megamap in the combinatorial mode linearly amplifies the difference in input strengths (Fig. 3(a) and (c)). In our simulations with \(N\approx 10\mbox{,}000\) place cells, the transition between modes occurs when the learned region reaches about 25 m^{2} [10].
The combinatorial mode is not commonly observed in attractor network models. Standard continuous attractor network models of place cells operate exclusively in the WTA mode unless the dynamical system is modified to make multipeaked activity bumps more stable [6, 37, 38]. It is interesting that the optimal megamap operates in either mode without any changes to the parameters or dynamical system, but the megamap with Hebbian weights operates in the WTA mode regardless of the environmental size. The emergence of the combinatorial mode not only depends on the environmental size but also on the manner in which the recurrent connections are updated as the animal explores novel regions of the environment.
2.3 Numerical Test for the Operational Mode
We now propose a numerical test for determining the operational mode of the dynamical system (Eq. (1)). We specify that the system is in the combinatorial mode if there exist stable fixed points with multiple activity bumps, and the network is in the WTA mode if any stable fixed point has exactly one activity bump.
In all numerical simulations we performed, the test is accurate in distinguishing between the two operational modes. For the example presented in Fig. 1, the recurrent weight matrix W is updated as the animal gradually learns novel subregions of an environment [10]. For the optimal weights, the test predicts the transition from the WTA mode to the combinatorial mode as the area (A) of the learned environment grows. In particular, \(r(\overline {S_{1}}\cup \overline {S_{2}},{\{\mathrm{inh}\} })\) decreases as A becomes larger, dropping below 1 around 25 m^{2} (Fig. 1(b), black closed circles). As predicted, when \(A<25\mbox{ m}^{2}\), exactly one activity bump persists in time given any initial state and any external input (Fig. 1(c)). When \(A>25\mbox{ m}^{2}\), two activity bumps persist in time given a mixed external input (Fig. 1(d)). For the Hebbian weights, the test predicts that the system remains in the WTA mode regardless of A since \(r( \overline {S_{1}}\cup \overline {S_{2}},{\{\mathrm{inh}\} })\) gradually increases with A (Fig. 1(b), gray closed circles). As predicted, we find numerically that two activity bumps are always unstable given Hebbian weights [10].
Equation (3) can also be used to test the stability of singlepeaked fixed points. Regardless of A or the method used to set the weights, \(r(\overline {S_{k}},{\{\mathrm{inh}\} })<1\) for any location \(\textbf {x}_{k}\) (Fig. 1(b), open circles and diamonds)). This indicates that any singlepeaked fixed point proportional to an embedded activity bump is stable. It is important to note that even in the combinatorial mode, the system robustly represents any location through a stable singlepeaked activity bump given a singlepeaked external input that may be relatively weak, noisy, or incomplete.
The numerical test is a powerful tool for determining the behavior of the network a priori. In addition to determining whether it is possible for multiple activity bumps to persist in time, the test determines whether the network may show hysteresis or amplify the difference in input strengths (Fig. 3(a) and (c)). However, the numerical test is limited in that it determines the stability but not the existence of a fixed point. Figure 1(b), open circles and diamonds, indicates that singlepeaked activity bumps are stable for any size environment. In our numerical simulations, we found that these singlepeaked fixed points always exist given the optimal weights, but all cells eventually become active when \(A=625\mbox{ m}^{2}\) given Hebbian weights [10]. Some sort of normalization, such as forcing the 1norm (subtractive normalization) or 2norm (multiplicative normalization) of the weight vector to be constant, would be required to maintain stability in the Hebbian network [33]. It would be interesting to examine in future work how normalization would affect the operational mode of the Hebbian network.
3 2Unit Reduced Model
While the numerical test of Eq. (3) can be used to determine the operational mode, we seek a deeper understanding of why the operational mode emerges in large environments, and under what set of parameters. We begin by reducing the model to a simple 2unit model that has similar dynamics and we can fully analyze.
3.1 Reduction of the Megamap Model to the 2Unit Model
Consider an external input that is some mixture of the two training inputs, \(\overline {\textbf{b}}(\textbf{x}_{1})\) and \(\overline {\textbf{b}}(\textbf{x}_{2})\) (Eq. (2)), where \(\textbf{x}_{1}\) and \(\textbf{x}_{2}\) are two wellseparated locations in the environment. We seek a mapping from the full megamap model to a twodimensional reduced model with the same form and the same qualitative dynamics given this conflicting external input. The simplest relevant simplification is to model two units, where the place cells in each unit k are given by the set \(\overline {S_{k}}\) (Eq. (4)), and the reduced state \(\widehat {u}_{k}\) is the collective state of place cells in unit k. The reduced model does not include cells without a place field near \(\textbf{x}_{1}\) or \(\textbf{x}_{2}\), as these cells should be silent (\(f(u_{i})\approx0\)) if the system is stable.
Reduced Model
Approximations in the Reduction
As detailed in Appendix B, we make four approximations to map the Ndimensional system of Eq. (1) to the twodimensional system of Eq. (5). First, we neglect cells that are in both units by assuming \(\overline {S_{1}}\cap \overline {S_{2}}= \emptyset\). Since place fields are set by the Poisson distribution, a small minority of cells in \(\overline {S_{1}}\) may also be in \(\overline {S_{2}}\), but these relatively few cells should not have a large impact on the dynamics. Second, we neglect the small minority of cells with multiple place fields near \(\textbf{x}_{k}\). This permits the assumptions that both units have the same number of cells, or \(\overline {N}= \overline {S_{k}}\) for any k, and that the average of the recurrent input (proportional to \(w^{0}\)) between two cells in the same unit given the embedded activity bump is the same for all k. Third, we neglect the asymmetries in the optimal weights of the megamap by assuming that the average weight from unit 1 to unit 2 (proportional to q) is the same as the average weight from unit 2 to unit 1. These first three approximations amount to neglecting the variability of the megamap and modeling only the average dynamics. The variability may affect the stability of a state in borderline cases. For example, when \(r( \overline {S_{1}}\cup \overline {S_{2}},{\{\mathrm{inh}\} }) \approx1\), the stability of two coactive bumps may depend on the locations chosen for \(\textbf{x}_{1}\) and \(\textbf{x}_{2}\).
3.2 Constraints on the Parameters of the 2Unit Model
 1.
The inhibitory unit must be active given a desired activity pattern, but inactive if all place cells are inactive. Equivalently, \(0<\theta<1\).
 2.
The strength of the training input must be much weaker than the desired equilibrium state, or \(0<\widehat {b}_{\mathrm{pk}}\ll1\). By Eq. (9), this condition is equivalent to \(\widehat {w}^{\mathrm {I}}(1\theta) \ll w^{0} < 1+\widehat {w}^{\mathrm{I}}(1\theta)\).
 3.When \(q=0\), the attractor of the megamap should consist of singlepeaked activity bumps. In the 2unit model, this means that when \(q=0\) and \(\widehat {\textbf{b}}=\textbf{0}\), the system supports fixed points in which exactly one unit is active. Without loss of generality, suppose that the fixed point in the absence of external input is given by \(\widehat {u}_{1}>0\) and \(\widehat {u}_{2}<0\). We show in Appendix C.2 that the inhibitory unit must be active at such a fixed point. By Eq. (5),Thus, this condition imposes the constraint, \(w^{0}>1\).$$ \begin{bmatrix} \widehat {u}_{1}\\ \widehat {u}_{2} \end{bmatrix} = \begin{bmatrix}w^{0}&0\\ 0&w^{0} \end{bmatrix} \begin{bmatrix} \widehat {u}_{1}\\ 0 \end{bmatrix}  \widehat {w}^{\mathrm{I}}( \widehat {u}_{1}\theta) \begin{bmatrix}1\\1 \end{bmatrix} \quad \Rightarrow\quad w^{0}1 = \widehat {u}_{2}/\widehat {u}_{1}. $$
 4.
Finally, the crossexcitation must be small enough such that the desired activity pattern is a fixed point of the system given the training input. With \(\widehat {b}_{1}=\widehat {b}_{\mathrm{pk}}\) and \(\widehat {b}_{2} = 0\), the fixed point must satisfy \(\widehat {u}_{1}=1\) and \(\widehat {u}_{2} = q \widehat {w}^{\mathrm {I}}(1\theta)<0\). Thus, this condition imposes the constraint, \(q<\widehat {w}^{\mathrm{I}}(1\theta) \ll w^{0}\).
4 Analysis of the Operational Modes of the 2Unit Model
In accordance with the definitions of the operational modes of the megamap, we specify that the 2unit model is in the combinatorial mode if there exist stable fixed points in which both units are active and in the WTA mode if any stable fixed point has exactly one active unit. We now analyze the 2unit model to derive an explicit equation for the critical value of \(w^{0}q\) at which the system shifts from the WTA mode to the combinatorial mode. We also analyze how the system responds to conflicting inputs in each mode, dependent on the attractor network strength (\(w^{0}q\)) and the relative strengths of the competing inputs (\(\widehat {b}_{1} \widehat {b}_{2}\)).
4.1 Characterization of the Operational Modes

At least one unit must be active at any stable fixed point due to the constraint, \(w^{0}>1\).
 A fixed point in which only unit 1 is active exists if and only ifSince \(w^{0}1<\widehat {w}^{\mathrm{I}}\), this fixed point exists for all inputs such that \(\widehat {b}_{1}\geq \widehat {b}_{2}\) if and only if \(w^{0}q>1\). If the fixed point exists, it is always stable and corresponds to the network encoding only the location with the stronger external input (\(\textbf {x}_{1}\)). The network effectively ignores the weaker input over location \(\textbf{x}_{2}\).$$ q < \bigl(w^{0}1\bigr) + \frac{( \widehat {b}_{1} \widehat {b}_{2})(\widehat {w}^{\mathrm {I}}(w^{0}1))}{\widehat {w}^{\mathrm{I}}\theta+ \widehat {b}_{1}}. $$(10)
 A fixed point in which only unit 2 is active exists if and only ifThis fixed point exists for some input such that \(\widehat {b}_{1}\geq \widehat {b}_{2}\) if and only if \(w^{0}q>1\). If the fixed point exists, it is always stable and corresponds to the network encoding only the location with the weaker external input (\(\textbf{x}_{2}\)). The network effectively ignores the stronger input over location \(\textbf{x}_{1}\).$$ q < \bigl(w^{0}1\bigr)  \frac{( \widehat {b}_{1} \widehat {b}_{2})(\widehat {w}^{\mathrm {I}}(w^{0}1))}{\widehat {w}^{\mathrm{I}}\theta+ \widehat {b}_{2}}. $$(11)
 A fixed point in which both units are active is stable if and only if \(w^{0}q < 1\). When \(w^{0}q<1\), such a fixed point exists if and only ifand the fixed point is unique.$$ q > \bigl(w^{0}1\bigr) + \frac{( \widehat {b}_{1} \widehat {b}_{2})(\widehat {w}^{\mathrm {I}}(w^{0}1))}{\widehat {w}^{\mathrm{I}}\theta+ \widehat {b}_{1}}, $$(12)
Setting \(\widehat {b}_{1}= \widehat {b}_{2}\) in (Eq. (12)), we conclude that the system is in the WTA mode when \(w^{0}q>1\) and in the combinatorial mode when \(w^{0}q<1\). This result is consistent with the hypothesis that the shift in operational mode observed in the megamap is due to the increase in crossexcitation between cells in the two respective activity bumps (Fig. 2(e)). Although the inhibitory weight and threshold (\(w^{\mathrm{I}}\) and θ, respectively) were not varied in our simulations of the megamap, the analysis of the 2unit reduced model implies that the operational mode depends only on the difference in self and crossexcitation, \(w^{0}q\), and not on \(w^{\mathrm{I}}\) or θ. This is somewhat surprising since the competition between two activity bumps, which underlies the WTA mode, is mediated by feedback inhibition.
In the WTA mode of the 2unit model, any stable fixed point represents exactly one location. This corresponds to the singlepeaked activity bumps always observed in equilibrium states of a relatively small megamap (Fig. 1(c), Fig. 3(a) and (c)). Since Eq. (10) is always satisfied, there are two stable fixed points for a given set of inputs (\(\widehat {b}_{1}\geq \widehat {b}_{2}\)) if and only if Eq. (11) is satisfied. In this case, the equilibrium state reached depends on the initial state, consistent with the hysteresis observed in the WTA mode of the megamap model (Fig. 3(c)).
4.2 Bifurcations of the Dynamical System

Type I: The state vector converges to a unique equilibrium in which only unit 1 is active.

Type II: The state vector converges to a unique equilibrium in which only unit 2 is active.

Type III: The state vector converges to one of two possible equilibria, one in which only unit 1 is active and one in which only unit 2 is active.

Type IV: The state vector converges to a unique equilibrium in which both units are active.
5 Conclusions
We present a mathematical analysis of the properties of the megamap attractor neural network that emerge when the network represents a sufficiently large spatial environment [10]. Through stability analysis of the full megamap model, we derive a numerical test (Eq. (3)) for determining the operational mode of the dynamical system (Eq. (1)). In addition, we derive a linear mapping from the Ndimensional megamap model to a twodimensional reduced model that has the same qualitative dynamics. Our analysis of the 2unit model elucidates the role of each parameter in the full megamap model in the context of conflicting external inputs (Fig. 4). In particular, we show that the abrupt shift in operational mode occurs when \(q \approx w^{0}1\), where \(w^{0}\) and q are proportional to the average recurrent excitation between two cells in the same unit and in different units, respectively (Eq. (7)). The inhibitory weight does not affect the operational mode, but increasing \(w^{\mathrm{I}}\) increases the range of q, resulting in a larger range of the relative strength of inputs (\(b_{\mathrm{pk}}^{1}b_{\mathrm{pk}}^{2}\)) for which there are two costable activity bumps (Type IV dynamics). The inhibitory threshold (θ) also does not affect the operational mode, but the bifurcations described by Eqs. (14)–(16) approach linear functions of \(b_{\mathrm{pk}} ^{1}b_{\mathrm{pk}}^{2}\) as θ approaches 1.
This work is similar in nature to numerous theoretical studies of EI nets [39, 40]. In many of these studies, two populations of neurons are considered, where one population represents excitatory cells and the other inhibitory cells. The recurrent circuitry among inhibitory cells is often neglected, simplifying the analysis. We consider two populations of excitatory neurons, each with extensive recurrent circuitry, and a third population of inhibitory neurons. We simplify the dynamical system by lumping all inhibitory neurons into a single inhibitory unit under the assumption that all inhibitory cells are statistically identical since interneurons in the hippocampus do not appear to have strong spatial tuning [41, 42]. We also assume that the time constant of the inhibitory state is much smaller than that of excitatory cells, allowing us to approximate the inhibitory state as an instantaneous function of the excitatory activity vector. Without this simplification, it is likely that we would observe oscillations between activity bumps under some parameter sets [18].
A common approach used to analyze continuous attractor neural networks is to approximate the Ndimensional system of ordinary differential equations (Eq. (1)) by a partial differential equation by taking the limit as \(N\rightarrow\infty\). The state vector, \(\textbf{u}(t)\in \mathbb{R}^{N}\), then becomes the continuous function, \(u(\textbf{x},t)\in \mathbb{R}\), where x is a continuous variable representing the single preferred location of a given place cell. The cleanest results are obtained using a Heaviside activation function for \(f(u)\), for then one can solve for the radius of the activity bump at a fixed point [14, 43]. Using a similar approach, we derived clean expressions for the set of stable fixed points; however, we found that the combinatorial mode does not exist given the Heaviside activation function in our dynamical system. Other mathematical studies have used Fourier analysis to analyze the PDE given the threshold linear activation function used for the megamap model [40, 44]. Even when we approximate the recurrent weights using only the first two terms in the Fourier series, however, the recurrent circuitry among both populations of neurons renders the solutions too complex to be helpful in understanding how the parameters of the model affect the dynamics. The approaches we present in this study require only a few justified approximations of the full megamap model, and the simplicity of the results make the analysis useful in understanding the behavior of the megamap. Despite its simplicity, the numerical test accurately determines the operational mode of the full system (Fig. 1), and the reduced model has similar qualitative behavior to the full model (Figs. 2 and 3).
While we focus on a particular attractor neural network, the results apply to a broad class of attractor network models. The numerical test for determining the operational mode (Eq. (3)) applies to any attractor network model in which the state vector is governed by Eq. (1), a standard firing rate model derived by averaging neuronal activity over multiple trials [33]. The reduced 2unit model applies to any attractor neural network in which the four approximations outlined in Sect. 3.1 are valid approximations. This includes not only continuous attractor neural networks, but also discrete attractor neural networks such as Hopfield networks with graded neuronal responses [2]. In the latter case, the set \(\overline {S_{k}}\) used in the reduction of the full model is the set of all cells that are active in embedded activity pattern k. It is not necessary for the embedded activity patterns to have the shape of the Gaussianlike activity bumps considered here.
When considering the reduced model, it is important to understand the impact of the approximations underlying the linear mapping from the full model. For the megamap, the first three approximations neglect the variability in embedded activity patterns and weights due to the Poisson distribution of place fields [10]. This variability includes asymmetries in the full weight matrix, W. We find numerically that, as long as W is a relatively small perturbation from a symmetric matrix, the asymmetries have a negligible effect on the dynamics. For example, we observe only a slight difference in the transition point between operational modes determined by numerical simulations and the stability test (∼25 m^{2}) and by the reduced state variables (\(w^{0}q\approx1.05\) at 25 m^{2}, as seen in Fig. 2(e)). This result is not surprising, as uncorrelated random perturbations of the weight matrix in a Hopfield network have been shown to have a small effect on the dynamics [45, 46]. The fourth approximation underlies the qualitative differences between the full megamap model and the 2unit model. In particular, the variable radius of the activity bump underlies the nonlinearities observed in the megamap’s response to the conflicting input (Fig. 3(a) and (c)). In general, the reduced model captures the peak of the activity pattern, but it does not capture changes in the subset of active cells within each unit.
A second possible extension would be to relax the fourth approximation of the reduced model to examine the spatial effects of the activity bump on the attractor. This could be done by modeling \(n\ll N\) place cells for each unit, setting the reduced weight matrix \(\textbf {W}^{0}\in \mathbb{R}^{n\times n}\) through a Gaussian tuning curve, and setting the reduced weight matrix \(\textbf{Q}\in \mathbb{R}^{n\times n}\) as a random matrix with \(\\textbf{Q}\\ll\\textbf{W}^{0}\\). It would be interesting to compare the operational modes and bifurcations of this 2ndimensional model to the operational modes and bifurcations of the twodimensional model presented here.
A third possible extension would be to use the reduced model to explore remapping. In the current study, the full weight matrix is set during a learning phase in which the place cell activity is fixed at the desired activity pattern, and the network is driven by strong external inputs. Then the dynamics of the model are examined during a retrieval phase in which the weights are constant, and the recurrent input is stronger than the external input. This separation into a learning phase and retrieval phase is common in attractor neural network models in which the weights are incrementally learned [6, 35, 47], and there is experimental evidence supporting, at least in part, the use of two separate phases. For example, it has been observed experimentally that the acetylcholine system is more activated during the initial exploration of a novel space than when the animal is moving around in a familiar space, and acetylcholine may increase the strength of afferent input connections relative to feedback recurrent connections [48]. Nonetheless, it would be an interesting and relevant study to address how the dynamics change given plasticity in the recurrent weights during the retrieval phase, as is more biologically realistic. Exploring remapping mathematically would require a more complex reduced model that incorporates differential equations for \(w^{0}(t)\) and \(q(t)\). The basic Hebbian learning rule is unstable, and the manner in which stability is maintained would affect the set of stable fixed points [33]. Another key factor would be the learning rate. In particular, when the two external inputs have equal strength, then two activity bumps initially become coactive in the WTA mode when the weights are constant. In the full model, this coactivity could last for hundreds of ms before one activity bump dominates [10]. Given Hebbian learning, the place cells in each unit would begin to reinforce each other’s activity, effectively increasing q and possibly driving the system to the combinatorial mode.
There are many contexts in which an attractor neural network must resolve conflicting information from its rich array of neuronal inputs. For example, it is a common experimental paradigm to manipulate different cues in different ways in order to track how information flows through various levels of neural processing [49, 50]. The WTA mode is ideal for robust memory retrieval, allowing the attractor network to perform computations such as transforming a noisy external input into a coherent, embedded activity pattern. On the other hand, the combinatorial mode permits a flexible recombination of embedded activity patterns in response to a changed environment. This flexibility could lead to phenomena such as the partial remapping observed in hippocampal place cells [6, 10, 31]. Perhaps the ideal attractor neural network operates between these two extremes, robustly encoding memories while still having the flexibility to adapt to our everchanging world. The reduction method presented in this paper is a useful tool for simplifying the mathematical analysis of various behaviors of attractor network models to better understand how these behaviors depend on the network parameters and the learning process.
Notes
Acknowledgements
Not applicable.
Availability of data and materials
Please contact author for data requests.
Authors’ contributions
KH conceived of the study, performed all analysis and simulations, and drafted the manuscript. KZ participated in the interpretation of results and in drafting the manuscript. All authors read and approved the final manuscript.
Funding
This work was funded by the Air Force Office of Scientific Research Grant FA95501210018 and by the National Institute of Mental Health Grant R01MH079511. Neither funding body contributed to the design of the study or the collection, analysis, or interpretation of data or in writing the manuscript.
Ethics approval and consent to participate
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Consent for publication
Not applicable.
References
 1.Amit DJ. Modeling brain function: the world of attractor neural networks. Cambridge: Cambridge University Press; 1989. CrossRefMATHGoogle Scholar
 2.Hopfield JJ. Neurons with graded response have collective computational properties like those of twostate neurons. Proc Natl Acad Sci USA. 1984;81:3088–92. CrossRefMATHGoogle Scholar
 3.Zhang K. Representation of spatial orientation by the intrinsic dynamics of the headdirection cell ensemble: a theory. J Neurosci. 1996;16:2112–26. CrossRefGoogle Scholar
 4.Redish AD, Elga AN, Touretzky DS. A coupled attractor model of the rodent head direction system. Netw Comput Neural Syst. 1996;7:671–85. CrossRefMATHGoogle Scholar
 5.Samsonovich A, McNaughton B. Path integration and cognitive mapping in a continuous attractor neural network model. J Neurosci. 1997;17:5900–20. CrossRefGoogle Scholar
 6.Stringer SM, Rolls ET, Trappenberg TP. Selforganising continuous attractor networks with multiple activity packets, and the representation of space. Neural Netw. 2004;17:5–27. CrossRefMATHGoogle Scholar
 7.Burak Y, Fiete IR. Accurate path integration in continuous attractor network models of grid cells. PLoS Comput Biol. 2009;5:e1000291. https://doi.org/10.1371/journal.pcbi.1000291. MathSciNetCrossRefGoogle Scholar
 8.Yoon KY, Buice MA, Barry C, Hayman R, Burgess N, Fiete IR. Specific evidence of lowdimensional continuous attractor dynamics in grid cells. Nat Neurosci. 2013;16:1077–84. CrossRefGoogle Scholar
 9.Cerasti E, Treves A. The spatial representations acquired in CA3 by selforganizing recurrent connections. Front Cell Neurosci. 2013;7:112. https://doi.org/10.3389/fncel.2013.00112. CrossRefGoogle Scholar
 10.Hedrick KR, Zhang K. Megamap: flexible representation of a large space embedded with nonspatial information by a hippocampal attractor network. J Neurophysiol. 2016;116:868–91. CrossRefGoogle Scholar
 11.Marr D. Simple memory: a theory for archicortex. Philos Trans R Soc Lond B, Biol Sci. 1971;262:23–81. CrossRefGoogle Scholar
 12.McNaughton B, Nadel L. Hebb–Marr networks and the neurobiological representation of action in space. In: Gluck MA, Rumelhart DE, editors. Neuroscience and connectionist theory. Hillsdale: Erlbaum; 1990. p. 1–63. Google Scholar
 13.Knierim JJ, Zhang K. Attractor dynamics of spatially correlated neural activity in the limbic system. Annu Rev Neurosci. 2012;35:267–86. CrossRefGoogle Scholar
 14.Amari S. Dynamics of pattern formation in lateralinhibition type neural fields. Biol Cybern. 1977;27:77–87. MathSciNetCrossRefMATHGoogle Scholar
 15.Ermentrout B. Neural nets as spatiotemporal pattern forming systems. Rep Prog Phys. 1998;61:353–430. CrossRefGoogle Scholar
 16.Veltz R, Faugeras O. Local/global analysis of the stationary solutions of some neural field equations. SIAM J Appl Dyn Syst. 2010;9:954–98. MathSciNetCrossRefMATHGoogle Scholar
 17.Stella F, Cerasti E, Treves A. Unveiling the metric structure of internal representations of space. Front Neural Circuits. 2013;7:81. https://doi.org/10.3389/fncir.2013.00081. CrossRefGoogle Scholar
 18.Li Z, Dayan P. Computational differences between asymmetrical and symmetrical networks. Netw Comput Neural Syst. 1999;10:59–77. CrossRefMATHGoogle Scholar
 19.Carroll S, Josić K, Kilpatrick ZP. Encoding certainty in bump attractors. J Comput Neurosci. 2014;37:29–48. MathSciNetCrossRefMATHGoogle Scholar
 20.Scoville WB, Milner B. Loss of recent memory after bilateral hippocampal lesions. J Neurol Neurosurg Psychiatry. 1957;20:11–21. CrossRefGoogle Scholar
 21.O’Keefe J, Nadel L. The hippocampus as a cognitive map. Oxford: Clarendon Press; 1978. Google Scholar
 22.Eichenbaum H, Cohen NJ. Can we reconcile the declarative memory and spatial navigation views on hippocampal function? Neuron. 2014;83:764–70. CrossRefGoogle Scholar
 23.Muller R. A quarter of a century of place cells. Neuron. 1996;17:979–90. CrossRefGoogle Scholar
 24.Fenton AA, Kao HY, Neymotin SA, Olypher A, Vayntrub Y, Lytton WW, Ludvig N. Unmasking the CA1 ensemble place code by exposures to small and large environments: more place cells and multiple, irregularly arranged, and expanded place fields in the larger space. J Neurosci. 2008;28:11250–62. CrossRefGoogle Scholar
 25.Park E, Dvorak D, Fenton AA. Ensemble place codes in hippocampus: CA1, CA3, and dentate gyrus place cells have multiple place fields in large environments. PLoS ONE. 2011;6:e22349. CrossRefGoogle Scholar
 26.Rich PD, Liaw HP, Lee AK. Large environments reveal the statistical structure governing hippocampal representations. Science. 2014;345:814–7. CrossRefGoogle Scholar
 27.Johnston D, Amaral DG. Hippocampus. In: Shepherd G, editor. The synaptic organization of the brain. New York: Oxford University Press; 1998. p. 417–58. Google Scholar
 28.Bliss TVP, Collingridge GL. A synaptic model of memory: longterm potentiation in the hippocampus. Nature. 1993;361:31–9. CrossRefGoogle Scholar
 29.Vazdarjanova A, Guzowski JF. Differences in hippocampal neuronal population responses to modifications of an environmnetal context: evidence for distinct, yet complementary, functions of CA3 and CA1 ensembles. J Neurosci. 2004;24:6489–96. CrossRefGoogle Scholar
 30.Rolls ET. An attractor network in the hippocampus: theory and neurophysiology. Learn Mem. 2007;14:714–31. CrossRefGoogle Scholar
 31.Colgin LL, Moser EI, Moser MB. Understanding memory through hippocampal remapping. Trends Neurosci. 2008;31:469–77. CrossRefGoogle Scholar
 32.Hebb DO. The organization of behavior: a neuropsychological theory. New York: Wiley; 1949. Google Scholar
 33.Dayan P, Abbott L. Theoretical neuroscience: computational and mathematical modeling of neural systems. Cambridge: MIT Press; 2001. p. 163. MATHGoogle Scholar
 34.Kali S, Dayan P. The involvement of recurrent connections in area CA3 in establishing the properties of place fields: a model. J Neurosci. 2000;20:7463–77. CrossRefGoogle Scholar
 35.Rolls ET, Stringer SM, Trappenberg TP. A unified model of spatial and episodic memory. Proc R Soc Lond B, Biol Sci. 2002;269:1087–93. CrossRefGoogle Scholar
 36.Solstad T, Yousif HN, Sejnowski TJ. Place cell rate remapping by CA3 recurrent collaterals. PLoS Comput Biol. 2014;10:e1003648. https://doi.org/10.1371/journal.pcbi.1003648. CrossRefGoogle Scholar
 37.Samsonovich A. Hierarchical multichart model of the hippocampal cognitive map. In: Proceedings of the fifth joint symposium on neural computation. 1998. p. 140–7. Google Scholar
 38.Moldakarimov S, Rollenhagen JE, Olson CR, Chow CC. Competitive dynamics in cortical responses to visual stimuli. J Neurophysiol. 2005;94:3388–96. CrossRefGoogle Scholar
 39.Wilson HR, Cowan JD. Excitatory and inhibitory interactions in localized populations of model neurons. Biophys J. 1972;12:1–24. CrossRefGoogle Scholar
 40.Hansel D, Sompolinsky H. Modeling feature selectivity in local cortical circuits. In: Koch C, Segev I, editors. Methods in neuronal modeling: from ions to networks. Cambridge: MIT Press; 1998. p. 499–568. Google Scholar
 41.Kubie JL, Muller RU, Bostock E. Spatial firing properties of hippocampal theta cells. J Neurosci. 1990;10:1110–23. CrossRefGoogle Scholar
 42.Wilson MA, McNaughton BL. Dynamics of the hippocampal ensemble code for space. Science. 1993;261:1055–8. CrossRefGoogle Scholar
 43.Ermentrout GB, Terman DH. Mathematical foundations of neuroscience. New York: Springer; 2010. CrossRefMATHGoogle Scholar
 44.BenYishai R, BarOr RL, Sompolinsky H. Theory of orientation tuning in visual cortex. Proc Natl Acad Sci USA. 1995;92:3844–8. CrossRefGoogle Scholar
 45.Hertz JA, Grinstein G, Solla SA. Memory networks with asymmetric bonds. AIP Conf Proc. 1986;151:212–8. CrossRefGoogle Scholar
 46.Crisanti A, Sompolinsky H. Dynamics of spin systems with randomly asymmetric bonds: Langevin dynamics and a spherical model. Phys Rev A. 1987;36:4922–39. MathSciNetCrossRefGoogle Scholar
 47.Widloski J, Fiete IR. A model of grid cell development through spatial exploration and spike timedependent plasticity. Neuron. 2014;83:481–95. CrossRefGoogle Scholar
 48.Hasselmo ME. The role of acetylcholine in learning and memory. Curr Opin Neurobiol. 2006;16:710–5. CrossRefGoogle Scholar
 49.Knill DC, Pouget A. The Bayesian brain: the role of uncertainty in neural coding and computation. Trends Neurosci. 2004;27:712–9. CrossRefGoogle Scholar
 50.Knierim JJ, Neunuebel JP. Tracking the flow of hippocampal computation: pattern separation, pattern completion, and attractor dynamics. Neurobiol Learn Mem. 2016;129:38–49. CrossRefGoogle Scholar
Copyright information
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.