Navigation and the efficiency of spatial coding: insights from closed-loop simulations

Spatial learning is critical for survival and its underlying neuronal mechanisms have been studied extensively. These studies have revealed a wealth of information about the neural representations of space, such as place cells and boundary cells. While many studies have focused on how these representations emerge in the brain, their functional role in driving spatial learning and navigation has received much less attention. We extended an existing computational modeling tool-chain to study the functional role of spatial representations using closed-loop simulations of spatial learning. At the heart of the model agent was a spiking neural network that formed a ring attractor. This network received inputs from place and boundary cells and the location of the activity bump in this network was the output. This output determined the movement directions of the agent. We found that the navigation performance depended on the parameters of the place cell input, such as their number, the place field sizes, and peak firing rate, as well as, unsurprisingly, the size of the goal zone. The dependence on the place cell parameters could be accounted for by just a single variable, the overlap index, but this dependence was nonmonotonic. By contrast, performance scaled monotonically with the Fisher information of the place cell population. Our results therefore demonstrate that efficiently encoding spatial information is critical for navigation performance. Supplementary Information The online version contains supplementary material available at 10.1007/s00429-023-02637-8.


Introduction
Spatial navigation is critical for the survival of most animals and many neural populations in the brain appear to have evolved to support spatial learning and navigation, e.g., place cells (PC) (O'Keefe and Nadel 1978), boundary cells (BC) Hartley et al. (2000), and head direction cells (Taube 1998).Place cells in the hippocampus are selectively active when the animal is located in one particular part of an environment, the place field (O' Keefe and Nadel 1978).A large number of studies have examined which aspect of the environment determines the firing patterns of place cells, e.g.there is a larger number of place fields with smaller sizes close to goal radius locations as compared to elsewhere in the environment (Parra-Barrero et al. 2021;Hollup et al. 2001;Lee et al. 2006;Ainge et al. 2007;Zaremba et al. 2017;Tryon et al. 2017;Gauthier and Tank 2018;Turi et al. 2019;Sato et al. 2020;Lee et al. 2020;Kaufman et al. 2020;Dupret et al. 2010;Grieves et al. 2016Grieves et al. , 2018;;Jarzebowski et al. 2022).The same observation was made near walls as compared to further away Tanni et al. (2022).Furthermore, the range of some place cell parameters can be large, e.g., place field sizes cover two orders of magnitude, ranging from tens of centimeters to more than 10 ms (Kjelstrup et al. 2008;Rich et al. 2014;Eliav et al. 2021).It is largely unknown why this variability occurs.
In sensory systems, it has been suggested that variability of the quality of sensory code might have an intrinsic source in the brain (Abbott and Dayan 1999;Rieke et al. 1999;Ralf and Bethge 2010).Barlow (1961) suggested the efficient coding hypothesis, which has driven much research in Behnam Ghazinouri and Mohammadreza Mohagheghi Nejad have contributed equally to this work.sensory neuroscience.A sensory code is considered efficient, if it maximizes the amount of useful information that can be extracted about a sensory input while minimizing noise and redundancy, or irrelevant information, (Graham and Field 2007;Wiskott and Sejnowski 2002).Many aspects of the efficient coding hypothesis have been tested and confirmed in sensory systems (Barlow 2001).Since place fields of hippocampal pyramidal neurons are similar to tuning curves in sensory neurons, the concept of efficient coding can readily be applied to spatial information, which we refer to as efficient spatial coding.One aspect of coding efficiency is how well place fields cover the environment, another is how many spikes place cells fire, and yet another is by how much the firing rate of place cells change as a function of spatial position.However, few studies have attempted to apply the efficient coding hypothesis to spatial representations, although there are notable exceptions (Rolls and Treves 2011;Mathis et al. 2012;Finkelstein et al. 2018).
Efficient coding would benefit spatial navigation similar to how it benefits perceptual tasks.However, the purpose of efficient sensory codes has not been studied nearly as much and remains obscure (Simoncelli 2003).Similarly, little is known about the exact functional role of place cells in spatial navigation and learning, in particular the role of place cell properties-such as field size, firing rate, and densityand their variations across the environment.At some point, it was hypothesized that place cells are important for path integration (Samsonovich and McNaughton 1997), which tracks the animal's position by integrating velocity signals.However, after the discovery of grid cells it was suggested that they might be a more likely candidate for maintaining a representation of the animal's position via path integration (McNaughton et al. 2006).Nevertheless, even if that were true, path integration is only one of the computations that is required for navigation behavior.So, even if spatial representations were intimately linked to path integration, it remains to be clarified how neural path integration impacts spatial behavior.
Here we studied the function of place cells in spatial navigation and the role of place cell parameters using an artificial agent based on spiking neural networks that learned to navigate to a goal location.This requires simulating neuronal processes and the emerging spatial behaviors a closedloop setup, in which neuronal activity is driven by the inputs from the environment, and which in turn affects actions.The action of the agent changes its state and the sensory inputs.We based our computational model on a computational modeling tool-chain (Jordan et al. 2019) that uses the NEST simulator (Gewaltig and Diesmann 2007) to simulate biological neural networks and Gym (Brockman et al. 2016) to model the interactions with the environment.In our simulation, the agent had to solve a spatial learning task similar to the Morris watermaze (Morris 1981).We found that, as expected, increasing the goal size leads to a better navigation performance.By contrast, the dependence of performance on the parameters of the place cell population, i.e., cell number, field size, and peak firing rate, was less clear.However, we found a systematic relationship between performance and a new variable, the overlap index, which measures the degree of overlap between two neighboring place cells and which incorporates both place cell number and field size.However, this relationship is nonmonotonic.Finally, we found that the Fisher information, which describes how informative the place cell population is about the spatial location (Kloosterman et al. 2014;Brunel and Nadal 1998), best accounts for navigation performance in our model.

Methods
For closed-loop simulations of spatial learning and memory we used the software framework CoBeL-spike (Closed-Loop Simulator of Complex Behavior and Learning Based on Spiking Neural Networks), which is available at https:// github.com/ sench eng/ CoBeL-spike (Fig. 1).CoBeL-spike builds on a computational modeling tool-chain (Jordan et al. 2019) that includes the NEST simulator for networks of spiking neurons (Gewaltig and Diesmann, M. 2007), such as place cells (O'Keefe and Nadel 1978), boundary cells Hartley et al. (2000), and head direction cells (Taube 1998).The network model is based on a earlier neural network developed by Brzosko et al. (2017) and described below.The interactions with the environment are modelled by OpenAI Gym (Brockman et al. 2016).

Overview of simulated task
We used CoBeL-spike to simulate spatial learning in a 2.4 m × 2.4 m open 2-d environment (which is comparable to sizes of Morris watermazes used in experiments).Each simulation consisted of 30 trials.At the beginning of each trial, the artificial agent was located at the center of an open environment.The agent moved freely inside the environment and sought a hidden goal zone, a circular area, centered at [0.5, 0.5] from the center of the environment.Its radius r varied across simulations, but remained fixed within a simulation.A trial ended either if the agent entered the goal zone or after the time out of 5 s (see below), whichever occurred first.The agent was rewarded, if it entered the goal zone.This reward drove learning, so that agents tended to reach the goal zone faster in later trials.Then the next trial commenced.
The main simulation parameters that were varied independently were the goal size r ∈ {0.1 m, 0.15 m, 0.2 m, 0.3 m} , the number of place cells in the input from N PC = 3 2 1 3 to N PC = 101 2 , and the place field size in the range of PC = 0.02 m to PC = 1.4 m.Varying the last two param- eters allowed us to study the role of the place cell population in spatial learning.For each parameter set, results were averaged over 100 simulations, in each of which the network was randomly initialized to simulate learning in different artificial agents.The number was chosen based on our experience that it reduced the variability of average performance to an acceptable level.
We used two variables to evaluate navigation performance of the agent in the simulated task: escape latency ( ) and hit rate.The former was determined by averaging the trial duration over the 100 repetitions of the same simulation.The lower , the better the navigation performance.The other performance variable was the hit rate, which was defined as the fraction of the successful trials out of the 100 repetitions of the same simulation.A trial was considered successful, if the agent found the reward within a certain time limit (see below).The higher the hit rate, the better the performance.Since both parameters measure performance, they are correlated to some extent, but that correlation can be high or low depending on the task and the behavior of the agents.For instance, if all agents learn the same trajectory to the reward, if they find it at all, then the latency and hit rate would be highly anti-correlated.This case is quite likely in simple task where there is one solution that is easily found and clearly better than other options.On the other hand, if there are multiple potential solutions that vary in their quality, it is conceivable that agents with the same success rate would show different latencies.
In preliminary simulations, we found that in most cases, the agent either found the reward within 3 s or did not find it at all.In most of the unsuccessful trials, the agent became stuck in a loop or failed to move.To make sure that simulation time was not wasted on these kinds of unsuccessful trials, we implemented a timeout after 5 s, which allows for some variation in the behavior on successful trials.If the trial duration was less than 4.5 s (there were very few data points between 4.5 s and 5 s), the trial was considered successful.Otherwise, the trial was considered unsuccessful.

Model architecture
In this section, we give an overview of the components in our computational model, which are described in more detail in Fig. 2. We performed our simulations with a time step of 0.1 ms.The agent's position x(t) modulated the fir- ing rate Ω i PC (x(t)) of N PC place cells (Fig. 2A), whose place field centers were evenly distributed across the environment (Fig. 2B).In addition, the input layer included eight boundary cells with firing rate Ω i BC (x(t)) (Fig. 2C).The output spikes of place and boundary cells were fed to 40 action selection neurons (Fig. 2D).The action selection neurons formed a ring attractor network.The location of the bump in this network determined the allocentric direction of the movement at each time step.A symmetric spike-timingdependent plasticity (STDP) enabled learning to the feedforward weights if the goal is found (Fig. 2E), mimicking the modulation of synaptic plasticity upon dopamine release (Brzosko et al. 2017) and the agent is not punished if it fails to obtain the reward.We utilized only positive reinforcement, similar to the Morris water maze paradigm (Morris 1981).

Spatial representations in the input layer
The The CoBeL-spike tool-chain consists of two main components: the environment and the artificial agent.The environment is simulated using OpenAI Gym (Brockman et al. 2016) and keeps track of the agent's position and reward and makes sure that the agent's moves are consistent with the physical constraints.The environment sends the current state and rewards to the learning agent and receives the action from the agent.The agent is a 2-layer spiking neural network implemented using NEST (Gewaltig and Diesmann 2007).Communication between two components is facilitated by two interfaces.MUSIC (Djurfeldt et al. 2010) transfers massive amounts of event information as well as continuous values from one parallel application to another.The action is taken through ZeroMQ (Hintjens 2013) to the Gym environment cells in our simulations were evenly distributed across the 2-d environment on a square mesh.The spiking activity of the i-th place cell is modeled as an inhomogeneous Poisson process, implemented as a Poisson generator (poisson_generator) in NEST (Gewaltig and Diesmann 2007).Thus, the spiking dynamics of place cells was not modelled.The firing rate of each place place cell was given by where is the coordinate of the place field center of the i-th cell, i PC is the maximum firing rate of that cell, and x,i PC and y,i PC are the standard deviations of the Gaussian kernel along the x-and y-axis, respectively.
We used a simplified version of Eq. 1, where the maximum firing rate of all cells are equal and the standard deviations of the Gaussian kernels all have the same value in both directions.
Hence, the firing map of the t-th place cell in our simulations can be simplified to Fig. 2A shows the firing rate map of two sample place cells with field sizes = 0.1 m and = 0.2 m.
Changing the place cell number N PC and field size PC changes the total spiking activity fed forward to the action selection neurons.This change causes a major impact on the agent's behavior.If the input is very low, the action selection neurons do not activate the ring attractor.If the input is too high, the ring attractor is saturated and there (1) is too much activity in the network and all action selection neurons become active.In both cases, a clear bump is not formed and the agent does not move properly.Our baseline for place cell number and place field size were N PC = 21 2 and PC = 0.2 m, respectively.Preliminary simulations showed that the best performance for these parameter values occurred when the maximum firing rate of a single neuron was PC = 200 Hz.In this case, the summed firing rate across all place cell at the center of the environment is 3500 Hz.Therefore, in all simulations we adjusted the maximum firing rate of individual cells such that the summed firing rate across all place cell at the center of the environment is 3500 Hz (see Fig. 2B), if possible.The maximum firing rate of individual cells so obtained are shown in Fig. 3.At first glance, these maximum firing rates might appear unrealistically high.However, for computational reasons our simplified model includes far fewer neurons than the hippocampus of rodents.We regard each neuron in our simulation as representative of a larger number of biological place cells that all have a place field in the same location.A rough back-of-the-envelop calculation shows that the total number of spikes that all place cells send to action selection neurons at a given moment in our model (3500 Hz) roughly matches the spike rate that downstream neurons might receive from hippocampal neurons.We assume that a neurons received projections from about 10,000 place cells, of which about 25% are active in a typical experimental environment.Assuming further that each place field covers about 15% of the environment's area yields about 375 neurons that are simultaneously active.If each neuron fires with a maximum firing rate of 10 Hz, the total input rate is about 3750 Hz.

A B C
Fig. 3 Scaling of maximum firing rate of place cells.To ensure that there was appropriate excitatory drive to the action selection network, we adjusted the maximum firing rate of place cells such that the summed firing rate at the center of the environment was 3500 Hz.Shown are the required maximum firing rates for three different simulation sets.A Variable place cell number N PC for fixed PC = 0.2 m.
We explore such different peak firing rates ( 8 ∼ 3500 Hz), because the firing rates of all fields at the center change when is adjusted.To make sure that the agent does not get stuck at the borders of the environment, we added eight boundary cells that allowed us to model the obstacle avoidance behavior of animals.Boundary cells were also modelled as inhomogeneous Poisson processes.The firing rate map of each boundary cell was rectangular of size 2x hw × 2y hw (hw stands for half width) centered at x i BC : The centers of four boundary cells are located in the middle of each border and the centers of four further border cells are located at each corner.The firing field maps of two sample boundary cells are shown in Fig. 2C.The boundary cells had a fixed projection pattern to the action selection neurons, such that the agent was repelled from the border or corner towards the inside of the environment.

Decision making with action selection neurons
At the core of the model network were 40 action selection neurons that formed a ring attractor.Each action selection neuron represents one preferred direction of movement j , distributed homogeneously over 360°.The cells were connected to each other though lateral synaptic weights such that neurons with similar preferred directions excite each other and neurons coding different movement directions inhibit each other (Brzosko et al. 2017), i.e., where w inh = −400 and w exc = 50 are inhibitory and excita- tory weights, respectively, is the Kronecker delta, and (3) = 20 .The lateral weights remained fixed throughout all simulations.
The parameter values ensured that the network is a continuous attractor network, where the attractor is an activity bump.That is, at any given point in time only a subset of neurons with similar movement directions was active and the rest of neurons were silent due to strong inhibition from the active neurons.
Feedforward inputs project from the place and boundary cells to action selection neurons.At the beginning of each simulation, the feedforward weights ( w ij fwd ) between input neurons and action selection neurons were randomly drawn from a normal distribution with mean 30 and standard deviation 5.The feedforward weights were plastic, and could be potentiated if the agent found the goal, up to a maximum of 60 (see below).
To visualize the postsynaptic effect of place cells and boundary cells onto action selection neurons, we represented each action selection neuron by a unit vector that points in the preferred direction of that neuron ( a j ) and then computed the sum over all such unit vectors weighted by the synaptic input that the neurons received from the input layer in a particular location (Fig. 2F).
We modelled neurons of the action selection network as leaky-integrate-and-fire' (LIF) neurons.The dynamics of sub-threshold membrane potential ( v i ) of the j-th LIF neurons is (Spreizer et al. 2019): where m = C m g L denotes the membrane time constant, C m the membrane capacitance, g L the leak conductance, E L the leak reversal potential, and I j (t) the total synaptic input current.The input current is the convolution of spikes and an alpha function (Rotter and Diesmann 1999;Bernard et al. 1994; (5)  Jack et al. 1983).The LIF neurons were implemented using the iaf_psc_alpha neuron in NEST (Gewaltig and Diesmann 2007;Rotter and Diesmann 1999;Diesmann et al. 2001;Kobayashi et al. 2009;Yamauchi et al. 2011).The values of the model parameters are summarized in Table 1.
To translate the activity of action selection neurons to navigation, the unit vectors a j representing the action neu- rons' preferred direction were weighted by the exponentially discounted spikes times and summed to obtain the movement vector selected by the network a(t) (Brzosko et al. 2017).
where s = 10 −4 m is the step size per spike and N AS = 40 the number of action selection neurons, and a = 0.5 ms.The second sum runs over all spikes that the action selection neuron j fired before the current time t.This scheme made it possible for the agent to move freely in any direction with different step sizes.
The activity of the action selection neurons was not changed between trials, i.e., the activity of the action selection neurons in the last moment of trial t is the same as at the beginning of the next trial t + 1 .Therefore, the initial conditions of trials are different from one another.This made it possible for the agent to find the goal after an unsuccessful trial, because otherwise the agent would repeat the same behavior on the next trial since weights do not change in/after unsuccessful trials.

Synaptic plasticity rules
The feedforward weights from place and boundary cells to action selection neurons play a critical role in our model.They define the policy of the agent.Plasticity of these weights enables the agent to learn the location of the goal zone.We applied a modified STDP rule that uses an all-to-all spike pairing scheme and eligibility trace that allows for delayed update.The following learning rule was applied to feedforward weights, if the agent finds the goal (Izhikevich 2007;Potjans et al. 2010).
where ẇ denotes the change applied to the feedforward weight connecting the i-th place cell to the j-th action selection neuron.c is an eligibility trace that allows for delayed updates, which is necessary and sufficient to learn causal ( 6) associations between synaptic activity when outcomes are delayed.This concept is a fundamental component of reinforcement learning network models (Brzosko et al. 2017;Izhikevich 2007;Legenstein et al. 2008;Florian 2007).n is the dopamine concentration, and b is the dopamine baseline concentration.c and n are the time constants of the eligibility trace and the dopamine concentration, respectively.(t) is the Dirac delta function, s pre∕post the time of a pre-or post-synaptic spike, s n the time of a dopamine spike.
The spiking relationship between pre-and post-synaptic neurons modulated plasticity via the following window function of additive STDP: where Δt = s post − s pre is the temporal difference between a post-and a pre-synaptic spike, A + the amplitude of the weight change, and + a time constant.We used the stdp_ dopamine_synapse in NEST (Gewaltig and Diesmann 2007) to implement this model with the parameters given in Table 2.

Characterizing spatial coding in the place cell population
Changing the number of place cells ( N PC ) and/or their field size ( PC ) changes the spatial coding in the place cell population.To quantify spatial coding we use two different measures.The first such measure, the overlap index, is defined by comparing two neighboring place fields, assuming that all fields are equidistant across the environment.So, the overlap index is the average firing rate of one cell Ω i−1 PC at the center of the next closest place field x i PC , normalized by the peak firing rate of the other neuron Ω i PC (x i PC ) .
(8) Note that by definition, the overlap index is independent of the peak firing rate of the individual neurons in our network.
We also quantified spatial coding by calculating the Fisher information matrix of the place cell population (Fisher 1922;Brunel and Nadal 1998): where q is the population activity vector.Given the place cell model, the Fisher information matrix at a location x is given by the following equation.
Since we use isotropic place fields, we average over the diagonal elements of the Fisher information matrix to quantify spatial coding by the place cell population.In the following, we refer to this average as the Fisher information of the place cell population.

Results
To study the functional role of the properties of place cells in spatial navigation, we performed closed-loop simulations of spatial learning controlled by a spiking neural network using the CoBeL-spike software framework.The behavior was similar to that of rodents in the Morris watermaze.Synaptic plasticity in the network based on STDP and an eligibility trace gradually changed the behavior of the agent such that it found the goal more often and faster in later trials (Fig. 4).
In other words, the agent successfully learned the spatial navigation task in our computational model.The outcome of learning is also evident in more direct trajectories to the goal and the policy encoded in the feedforward projections between spatial inputs and the action selection network (Fig. 5).In the following, we study systematically how task and place cell parameters affect navigation performance on this spatial learning task. (10) x 2 p(q|x)dq, Studying the impact of cell number, field size, and goal size on navigation performance First, we kept the field size of the place cells unchanged ( PC = 0.2 m), while we varied the place cell number from 3 2 to 101 2 ) and goal size.We adopted the place field size PC = 0.2 m.To ensure that the agent is initially located at the center of a place cell at the start of each trial, we consistently assigned an odd number to the place cell numbers in a row.This approach helps to drive the agent initially and maintains consistency across trials.The smallest practical number for place cell is therefore 3 2 .Our perfor- mance measures are escape latency (the lower, the better) and the hit rate (the higher, the better).We observed that the simulation results based on escape latency (Fig. 6A) and the hit rate (Fig. 6D) for this parameter set are consistent with one another.As mentioned above, this indicates that all agents that are successful find similar trajectories to the goal.Hence, we will only present the escape latency for any further computations or analyses.A clear result that we found throughout all simulations was that the agent performs better as the goal becomes larger.This occurs because the task becomes easier in two ways: the distance between start and goal becomes shorter and a less precise policy suffices to reach the goal zone.
By contrast, the relationship between performance and cell number was less clear.Increasing the cell number increases the density of place field, which in turn improves the spatial coverage by the place cell population.So, larger cell number should result in better navigation performance.We varied the cell number from 3 2 to 101 2 .We report cell number using this convention, because the cells are always located on a 2-d grid (Fig. 2B).The agent performed the worst for very small cell number and gradually performed better when cell number increased-as predicted (Fig. 6A,  D).However, for a large number of cells, the performance declined.We hypothesize that this occurs, because the place fields have too much redundancy, i.e. overlap.
Next, we changed the place field size while keeping the cell number constant at N PC = 21 2 (Fig. 6B, E).We chose Fig. 4 Learning curves in the simulated task are sensitive to network and task parameters.Examples of escape latencies (left column) and hit rates (right column) as a function of trial number for three sets of network and task parameters.For some parameter learning is evident in decreasing latency and increasing hit rate (red and green lines), whereas for others there is little to now learning (blue line) that value because the agent performed well in the previous simulation.We expected that the smaller the fields, the more precise the spatial coding and the better the behavioral performance will be.Hence, we reversed the x-axis for this plot from the standard convention (left: PC = 1.4 m, right: PC = 0.02 m).As expected, performance was better (lower and higher hit rate) for smaller place field sizes, but, like above, there was an optimal field size beyond which the performance declines.We hypothesize that very small fields cannot properly cover the environment, which leads to issues with navigation and/or spatial learning.

Investigating possible invariance to spatial coverage
Taken the results of the two simulation sets together, we hypothesize that the coverage of the environment by the place cell population might be the variable that determines navigation performance.We therefore defined the coverage index as follow: To test our hypothesis, we next changed the cell number and field size in such a way that leaves the coverage index constant at N PC 2 PC = 17.64 m 2 .This value arises from our standard parameters for N PC = 21 2 and = 0.2 m.Perfor- mance is roughly constant for a sizable range of place cell parameters (Fig. 6C, F), but at lower cell numbers and larger field sizes, the performance deviates quite dramatically.So, coverage alone is not sufficient to explain the difference in navigation performance.

Accounting for navigation performance using measures of spatial coding by the place cell population
We hypothesized further that it is not just the spatial coverage by place cells per se, but more subtle properties of spatial coding by the place cell population that matter for spatial navigation.We therefore defined the overlap index to measure spatial coding by quantifying how close the peaks of neighboring place fields are relative to their width.The overlap index increases if fields move closer together while their sizes PC remain constant.This occurs when only the cell number N PC is increased (Supplementary Fig. A1, from left to right).The overlap index also increases if field sizes increase while the field density stays constant, i.e., the cell number is constant (Supplementary Fig. A1, from bottom to top).We calculated the overlap index for the three sets of simulations that we performed above (Fig. 7, left column).Per definition, the overlap index increases from 0 to 1 with cell number (Fig. 7A), and decreases from 1 to 0 with decreasing field size (Fig. 7B).Also, as intended the overlap index deviates from the coverage index, i.e., when the coverage index is held constant while place cell parameters change, the overlap index varies (Fig. 7C) Interestingly, when we plotted navigation performance vs. the overlap index, a pattern emerged that is similar across the three simulation sets (Fig. 7, right column).However, the relationship is nonmonotomic (Fig. 7D, E).What is more, the asymmetric u-shape has a very wide bottom, so that the best performance is observed around a large range of values of the overlap index between 0.2 and 0.8 (Fig. 7D, E).These results are clear indications that the overlap index accounts for spatial coding features that are important for spatial navigation, but does not capture it well enough.
Next, we used a measure of sensory coding that is often used in sensory neuroscience (Brunel and Nadal 1998): the Fisher information.We calculated the Fischer information for the place cell population at all locations between the start point (center of the environment) and the center of the reward zone (see Sect. 2.6) and use the minimum value encountered to characterize the place coding by the population-more on this point below.The Fisher information increases if the number of cells increase while field sizes remain constant (Supplementary Fig. A2, from first to second column).Since the Fisher information also takes into account the peak firing rate, we held the peak firing rate constant for this comparison.However, if we scale the peak firing rate as we do in our closed-loop simulations, the simultaneous scaling of cell number and peak firing rate leaves the Fisher information virtually unchanged (Supplementary Fig. A2, from first to third column).When field sizes increase and peak firing rates simultaneously decrease, These properties of the Fisher information explain its scaling in the three simulation sets (Fig. 8, left column).When cell number is increased, and peak firing rate is decreased, while field sizes remain constant, the Fisher information remains constant (Fig. 8A).In the other two cases, the Fisher information increases as expected (Fig. 8B,  C).However, for very low cell numbers or very small field sizes this trend breaks, because in these cases the place fields are so sparse that the Fisher information varies substantially between locations where there is a place field and those where there is none (see Supplementary Fig. A3).This is the reason why we chose the minimum Fisher information along the path to quantify spatial coding in the place cell population.
The navigation performance vs. Fisher information plots are similar across all three simulation sets (Fig. 8, right column) and, the relationship is largely monotonic, unlike the relationship between performance and overlap index.We

Consistent relationship between navigation performance and efficiency of spatial coding
To investigate whether the relationship between navigation performance and spatial coding is the same in all three simulation sets, we plotted the results of all simulations in one figure.Both relationships navigation performance vs. overlap index (Fig. 9A) and navigation performance vs. Fisher information (Fig. 9B) are consistent across all three simulation sets.Our results therefore show that the efficiency of spatial coding of the place cell population is a determinant of the behavioral performance of the agent in the closed-loop simulations, regardless of the specific value of other place cell parameters.

Discussion
Using closed-loop simulations of spatial learning based on spiking neural networks and reward-driven STDP, we studied how the agent's navigation performance depends on the goal size and the parameters of the place cell population.While the dependence on goal size is straight forward, the larger the better, the dependence on cell number and field size is complex and difficult to interpret.We found that the Fisher information of the place cell population, which measures the efficiency of spatial coding, provided a simple correlate of navigation performance.Therefore, our work not only shows that the efficient coding hypothesis can be applied fruitfully to spatial representations, it also demonstrates that efficient spatial coding is important for function, i.e. spatial navigation and learning, as well.

Related closed-loop simulation work
The closest simulation framework to ours is the one that our model was based on Brzosko et al. (2017).That study used the model to study the role of sequential modulation of synaptic plasticity rule in learning a navigation task, but did not study the role of place cell parameters.Another closedloop simulator based on spiking neural networks, SPORE Kaiser et al. (2019), provides an interface between NEST (Gewaltig and Diesmann 2007) and the GAZEBO robotics simulator,1 but primarily targets robotics tasks.Furthermore, there is a framework for simulating closed-loop behavior based on reinforcement learning that is oriented towards neuroscience (CoBeL-RL) (Walther et al. 2021).While this allows for studying the computational solutions that emerge under different constraints, such as place-cell like representations Vijayabaskaran and Cheng (2022), how the statistics of hippocampal replay emerges in the network Diekmann and Cheng (2022), or the function of memory replay Zeng et al. (2022), these studies were based on machine learning implementations of reinforcement learning and neural A B Fig. 9 Summary of dependence of navigation performance on key parameters.Two variables are sufficient to account for task performance (as measured by escape latency) across all simulations: The reward zone radius (indicated by color) and spatial coding efficiency.A The dependence of latency on reward zone size and overlap index is consistent across all simulations.The best performance is achieved around overlap index= 0.5, regardless of goal size.B The dependence of latency on Fisher information is also consistent across all simulations, albeit somewhat more noisy.However, the relationship is monotonic and therefore the Fisher information captures a more fundamental property that is important for spatial navigation, i.e. the efficiency of spatial coding networks.Even though they help us understand the abstract computations performed by the brain during spatial navigation, the internal processes of this model significantly differ from those in the brain.In addition, there are simulation frameworks that target machine learning or industrial applications, such as, for example, DeepMind Lab (Beattie et al. 2016), RLlib (Liang et al. 2018), MiniGrid,2 and MAgents (Zheng et al. 2018).

Efficiency of spatial coding
Spatial coding by individual place cells has been quantified since the early 1990's by the spatial information.It is based on Shannon information and was introduced to hippocampal research by Skaggs et al. (1992).However, spatial information has only been used to quantify the degree of spatial modulation of one place cell's activity map, but not the spatial coding efficiency in the population.
The more detailed results aside, our results imply a simpler relationship to roughly maintain a similar level of coding efficiency in a limited parameter range: the smaller the fields are, the larger the number of fields should be to maintain the same level of coding efficiency.This principle accords well with experimental observations and provides an explanation for the observation that when smaller place fields are observed in an area of an environment, the density of fields there is higher too (Hollup et al. 2001;Lee et al. 2006Lee et al. , 2020;;Ainge et al. 2007;Zaremba et al. 2017;Tryon et al. 2017;Gauthier and Tank 2018;Turi et al. 2019;Sato et al. 2020;Kaufman et al. 2020;Dupret et al. 2010;Grieves et al. 2016Grieves et al. , 2018;;Jarzebowski et al. 2022;Tanni et al. 2022).

Limits of efficient coding
However, the observation that field sizes vary systematically with environmental features and behavior (Parra-Barrero et al. 2021), suggests that absolute field sizes do matter.If the Fisher information was the only variable that mattered for navigation performance, the hippocampus could as well have allocated a different number of fields with different sizes as long as the Fisher information is preserved.However, the brain consistently opts for more and smaller fields in certain situations, so this suggests that there are factors beyond the Fisher information that constrain spatial coding in the place cell population.Some variables that might impose additional constraints are the size of obstacles (our simulation did not include any), and temporal coding linked to the behavior of the animal, such as running speed (Parra-Barrero et al. 2021).In addition, there are more generic limitations on efficient coding, i.e. energy consumption (Crotty et al. 2012), and limits on peak firing rate of neurons due to spike dynamics.Future work is required to explore these factors and the trade-off between them.

Learning vs. navigation
The current study distinguishes between spatial navigation and spatial learning.Here, we define spatial navigation as the set of movements that an agent makes to reach certain locations in an environment.Spatial learning is the acquisition of spatial information and/or the process of determining which movements are most beneficial for the agent.In our model, the agent does the latter, i.e. it learns which actions to perform in certain spatial locations.This corresponds to what has been called a response strategy.In our model, there are limitations on spatial navigation because of the architecture of the neural network and the inputs it receives (from place cells).For instance, due to the finite number of action selection neurons and the exponential kernel in the output (Eq.6), the agent's movements have a limited temporal and spatial precision, i.e. they cannot be arbitrarily precise.Furthermore, there are limitations on spatial learning.Even if the network weights could, in principle, be set up such that it would perform optimal navigation, the synaptic learning rule and the statistics of learning might not be able to bring about such network weights.In other words, a solution might exist, but the learning process cannot find this rule.
To begin dissociating the limitations on learning from those on navigation itself, we compared the performance of trained networks to randomly initialized networks.We found that some difference in performance was already present in the random network before learning commenced (Supplementary Fig. A4), suggesting that for some place cell parameters, navigation is inherently more robust than for others and that learning amplifies this difference.However, it remains an open question whether other ways of optimizing synaptic weights, i.e., avoiding learning via reward-driven STDP, might yield a better solution for spatial navigation.In cases, where this is possible, the limitation of the network is primarily on learning, not navigation.In cases, where a wellperforming solution cannot be found by other means either, the limitation is on navigation per se.This requires a more systematic exploration of different optimization algorithms that is beyond the scope of this paper.

Conclusion
While the encoding of sensory information has received a great deal of attention in neuroscience, the functional role of sensory representations is much less clear.This applies especially to spatial representations and navigation.Here, we have shown using computational modeling that the efficiency of spatial coding in the inputs directly influences the navigation performance of an agent based on spiking neural networks and reward-modulated STDP.

Fig. 1
Fig.1Relationship between the components of the CoBeL-spike framework.The CoBeL-spike tool-chain consists of two main components: the environment and the artificial agent.The environment is simulated using OpenAI Gym(Brockman et al. 2016) and keeps track of the agent's position and reward and makes sure that the agent's moves are consistent with the physical constraints.The environment sends the current state and rewards to the learning agent

Fig. 2
Fig. 2 Illustration of model components.A Place cells of different field sizes.Left: firing maps of two CA1 place cells (Herzog et al. 2019).Firing rate is indicated by color (from blue to red).Right: two sample firing fields of place cells in the model (top: = 0.2 m and bottom: = 0.1 m).The red disk is the goal zone ( r = 0.2 m).Once the agent enters to this zone, the trial is terminated and the agent is rewarded as described in Sect.2.5.B Left: Firing rate maps of two individual place fields of different sizes at the center of the environment (top: = 0.2 m, bottom: = 0.1 m).Also shown is the distribution of place field centers of the other cells (top: N = 441 , bottom: N = 49 , marked with yellow dots).Right: The summed firing rate map over all cells.The maximum firing rate of individual cells is chosen such that the summed firing rate at the center of the environment is 3500 Hz.C Left: Firing rate maps of two boundary cells from MEC (van Wijngaarden et al. 2020).Right: two boundary cells from our model.D Illustration of the model architecture.Left: Place cells and boundary cells encode the position x(t) of the agent and provide

Fig. 5
Fig. 5 Example of behavior and learned policy during training.In the simulated task, the agent starts every trial in the center (green dot) and has to navigate to the goal zone (red disk) to receive reward.A-C Sample trajectories of the agent in the first (A), fifth (B), and tenth (C) trial.D-F Sum of the effects of all place cells on action selection

Fig. 6
Fig. 6 Asymptotic navigation performance after learning.A-C Escape latency as a function of place cell number N PC for fixed PC = 0.2 (A), of place field size PC for fixed N PC = 21 2 (B), and simultaneous scaling of both variables such that N PC × 2 PC = 17.46 (C).Each point represents the average trial durations over 100 repetitions.The color represents different goal radii (as indicated).D-F The hit rate, i.e. the fraction of trials in which the agent successfully navigates to the goal, shown using the same plotting convention as in (A-C)

Fig. 7 F
Fig.7The agent's navigation performance shows a dependence on overlap index.The dependence of the overlap index on key simulation parameters: A cell number when sizes are fixed, B field size when numbers are fixed, and C when cell number and field sizes scale inversely to one another.D-F Escape latency depends on overlap index for the three types of scaling, but the relationship is non-monotonic

Fig. 8
Fig.8Fisher information accounts for agent's navigation performance.Same plotting convention as in Fig.7for the binary logarithm of the minimum value Fisher information on the direct path between the start point and the goal's center.A Except for very low cell numbers, the Fisher information remains constant when increasing place cell number in our simulations, because we simultaneously decrease the maximum firing rate of individual cells (see Supplementary Fig.A2).B Decreasing the place field size lead to an increase in Fisher information, because of the power of 4 PC in the denominator of Fisher information.However, for extremely small place fields the Fisher information drops because the place representation becomes too sparse.C Scaling of cell number and field size simultaneously also increases Fisher information.D-F For the three types of parameter scaling, the escape latency monotonically decreases, and therefore the performance increases, with increasing Fisher information For higher values of , more fields contribute to the summed firing rate at the center.B Variable place field size PC for fixed N PC = 21 2 .
C: Both cell number and field size change simultaneously, such that N PC × 2 PC = 17.46

Table 2
STDP learning parameters, implemented to feed forward weights from place cells to action selection neurons