acceptodds
Under review as a conference paper at ICLR 2027

Are Grid Cells an Efficient Encoding of Space?

Abstract

Since their discovery over 20 years ago, grid cells have attracted considerable theoretical interest. Although it is widely agreed they subserve path integration (the tracking of position from self-motion cues), the precise reason for their distinctive activity remains debated. Previous work has generated grid-like representations by optimising objectives motivated from both the efficient coding of spatial position and path integration, often encouraging properties such as conformal isometry, translation-invariant dynamics or high coding capacity. These studies have identified important conditions under which grid-like codes emerge. However, whether grid cells emerge from directly optimising for efficient coding alone, across a broad class of possible representations, has not been fully addressed and still remains an open question. To address this question, we consider a natural task-level formulation of efficient coding: under biologically plausible constraints, we seek representations that minimise the mean squared error in the presence of noise. Specifically, we consider an encoder that receives noisy displacement updates and whose activity itself is corrupted by noise, deriving conditions under which this path-based problem reduces to the classical efficient-coding problem of encoding a source distribution. To analyse this problem, we employ the Ziv–Zakai lower bound, an information-theoretic tool that captures both local decoding precision and catastrophic errors arising from periodic spatial codes. Along the way, we see that several properties previously found to produce grid cells are theoretically advantageous within our coding framework. Grid-cell-like representations then emerge from the optimisation, providing evidence that grid cells constitute a highly efficient spatial code. More broadly, our results support the idea that the problem of efficient spatial coding is tightly related to the one of path integration, this suggests that task-relevant compression can favour representations that reflect the symmetries and compositional structure of the task.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Grid Cells, Theoretical Neuroscience, Efficient Coding, Information Theory, Normative Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wreKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explains grid-cell representations from an efficient-coding perspective, using the ZZLB to account for both local decoding errors and confusion between distant locations, and obtaining grid-like responses through optimization within a restricted Fourier encoding family. Its main strength is that an error lower bound derived within the efficient-coding framework serves as an optimization objective that unifies several computational principles previously proposed for obtaining grid cells through optimization. This helps simplify our understanding of the field.

My main concern, however, is the gap between the paper’s intended novelty and the evidence provided. Many previous studies have already discussed the efficiency of grid codes, so this paper needs to go further and ask whether grid representations would still win out because of their coding efficiency when a broader range of candidate encodings is allowed to compete. However, the authors ultimately restrict the search to a particular Fourier encoding family, without proving that this restriction does not exclude more efficient codes. The final result is also a numerical solution obtained through gradient-based optimization, rather than a theoretically derived optimal representation, and it is not even guaranteed to be globally optimal within this restricted family. This makes the experiments particularly important: broader comparisons under matched constraints are needed to demonstrate the efficiency advantage of grid representations. Simply showing that grid-like patterns can be obtained by optimizing within a selected encoding family provides limited additional insight beyond existing work.

The organization also needs substantial revision. Sections 2, 3, and 4 could be combined and shortened, with many definitions and standard derivations moved to the appendix, so that the central ideas and the advance over previous work become clear much earlier.

I would currently give this paper a 2. If the authors substantially improve the organization and provide more convincing experiments supporting the paper’s core novelty and coding-efficiency claims, I would consider raising it to a 3.

Critical strengths

I think the main strength of this paper is that it uses an error lower bound within the efficient-coding framework as an optimization objective, accounting for both local decoding errors and confusion between distant locations while connecting several computational principles used in previous work to obtain grid-cell representations through optimization. This provides a unified perspective that helps simplify our understanding of these principles and their relationships.

Critical weaknesses
  1. The current experiments mainly answer whether grid-like patterns can be obtained within the chosen encoding family, rather than adequately addressing why grid codes would be more efficient than other candidate encodings. This distinction matters because previous work has already discussed certain efficiency advantages of grid codes. The authors restrict the encoding form and then use gradient-based optimization to find a numerical solution. This neither rules out better performance from more general encodings nor guarantees that the best solution within the chosen family has been found. I would like to see stronger experimental evidence in two complementary directions: direct comparisons between different encoding approaches under matched constraints, and optimization of the same ZZLB objective with the non-negativity constraint in a more general encoding framework, removing the restriction to the prescribed Fourier basis. The latter would test whether grid-like representations still emerge when the search is allowed to explore a broader range of encodings.

  2. The paper takes too long to get to the point, and the central contribution gets buried in the mathematical setup. By the end of Sections 2–4, readers have mainly been given the task definition and several standard reductions, while the paper’s most important new insight remains difficult to identify. I suggest combining these sections into one, retaining only the setup, conditions, and conclusions needed to understand the rest of the paper, and moving the detailed derivations to the appendix. This would bring the central idea of unifying the coding principles into focus much earlier.

Questions for authors
  1. At the beginning of Section 3, the phrase “prove when” path-independent encoding is optimal led me to expect a further explanation of when path information can be ignored and when it remains important. What is currently provided is mainly a conditional-independence formulation and a Gaussian example in the appendix. Could you explain the scope of these conditions more clearly? I suspect that whether path details are useful may depend on the animal’s movement strategy, and I wonder whether this could be related to distortions in grid representations during goal-directed behavior, such as those reported by Boccara et al. (2019, Science).

  2. The Jensen argument in Eq. (19) favors spatial uniformity when the mean squared neural distance is fixed for each displacement magnitude and direction, but these means can themselves vary across encoders. I think translational symmetry is a reasonable structural constraint, but the distinction needs to be clearer: the earlier conditional result motivates spatial uniformity, whereas Section 6 imposes it as an exact constraint on the encoding form. Section 7 already acknowledges that this restriction may exclude more efficient codes. Could this be clarified when introducing Eq. (25), to avoid giving the impression that a lossless reduction follows from Jensen’s inequality?

1 comment
  • W1 "I would like to see stronger experimental evidence in two complementary directions: direct comparisons between different encoding approaches under matched constraints, and optimization of the same ZZLB objective with the non-negativity constraint in a more general encoding framework": We thank you for this suggestion and in doing so we believe our paper is now strengthened greatly. We hope we have managed to address both of your directions, through our new non parametric functional optimisation and new our grid vs place cell comparison, the full details and discussions of the results of which can be found in the restructured numerical experiments section.

  • W2 "The paper takes too long to get to the point" "Sections 2, 3, and 4 could be combined and shortened": We really appreciated this comment and completely agree with your assessment. We have implemented the changes as suggested and think the readability of the paper and clarity of the contributions is much improved. The main changes were combining section 2-4 as suggested which then allowed us more space to give intuition on the ZZLB including adding points touched on by reviewer 9SYc, explicitly writing the Jensen's gaps for clarity, a new numerical section better motivating the role and intention of each experiment alongside the 2 new experiments (which we detail in the general comment) and adding a contributions section.

  • Q1: "I wonder whether this could be related to distortions in grid representations during goal-directed behavior, such as those reported by Boccara et al. (2019, Science)" Thank you; this is an interesting connection. We see a more direct link through the source distribution than through path invariance. Our uniform source distribution, as mentioned in the paper represents, a large empty circular arena. Introducing rewards could instead motivate greater coding precision near these salient locations, modelled by assigning them greater weight, due to their increased perceived importance. The source distribution can therefore represent expected occupancy or, when interpreted as an effective weighting, the relative coding importance of different locations. As mentioned in the paper we tried to consider the most general simple model, as a next step alongside the consideration of boundaries this would be very interesting!

  • Q2 "The Jensen argument in Eq. (19) favors spatial uniformity when the mean squared neural distance is fixed for each displacement magnitude and direction, but these means can themselves vary across encoders." Following your organisational comments, we have clarified the intuition behind this term, particularly through the comparison with Dorrell et al. (2022). They constrain the neural update rule to depend only on source displacement, whereas our term encourages the resulting neural distance (which themselves are caused by the update rule, which we do not constrain) to depend only on displacement.

  • Q2 "the earlier conditional result motivates spatial uniformity, whereas Section 6 imposes it as an exact constraint on the encoding form" "Could this be clarified when introducing Eq. (25)" We agree this should have been made clear earlier, this has been updated, as well as us making the intentions of this experiment more clear.

Reviewer_EC5EOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper attempts to identify efficient neural codes for two-dimensional space using efficient-coding principles and whether grid-like neural representations would emerge from these principles alone. The study is motivated by the conflicting arguments existing in the literature that on the one hand say that grid cells do not provide the most efficient spatial code while on the other hand there are investigations on how grid-like representations arise in neural networks trained for path integration. It presents a theoretically grounded efficient-coding coding setup and experimentally validates the emergence of grid-like neural representations from the efficient-coding coding setup alone.

Critical strengths

The proposed efficient-coding coding setup for studying the problem imposes biologically plausible non-negativity and energy constraints.
Within the efficient-coding setup, the authors show that applying Ziv-Zakai lower bound allows capturing the thresholding effect (small local errors transit into large global errors) that is not captured by commonly used Cramér-Rao lower bound. Numerical results demonstrate the emergence of both hexagonal lattices and multiple modules of varying scales.

Critical weaknesses

None

Questions for authors

While this does not change the overall assessment of the paper, the reviewer would like to see in Discussion how the results reported in this paper allow to reconsolidate (and, perhaps, rethink) the conflicting findings that motivate the paper. The reviewer noticed a few typos while reading the paper: “castestrophic errors” -> “catastrophic errors” “gaussian” -> “Gaussian” equation 9 vs. Eq. 9 (and for other equations)

1 comment

We would like to thank you for the time taken to review our paper and the kind words. We have implemented the changes suggested, alongside further changes as suggested by Reviewers wreK and 9SYc, which we believe has allowed us to further improve the paper, we detail these changes in the general comment.

Reviewer_9SYcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether grid cells follow from efficient coding alone, without a path-integration objective. An animal integrates noisy self-motion cues, a population of 73 neurons encodes the result, channel noise corrupts the code, and a decoder recovers position with minimum mean squared error under non-negativity and an energy bound. For a Gaussian random walk the integrated cue is a sufficient statistic, a path-dependent encoder cannot do better, and the problem reduces to the coding of one position variable plus a constant drift term. The mean squared error is then replaced by an orientation-averaged Ziv–Zakai lower bound, relaxed with Jensen's inequality into four design principles, and minimised over a translation-covariant Fourier code. The optimised code shows hexagonal rate maps at three scales, from which the abstract concludes that the results provide "evidence that grid cells constitute a highly efficient spatial code".

Critical strengths

The theory seems correct where it was checked.

The choice of the Ziv–Zakai bound is interesting, and more than the paper says. Let us consider a Fisher-information objective on the same Fourier class. The Fisher information of Eq. 25 is Σ a_m² k_m k_mᵀ/σ_c², so the Cramér–Rao bound falls to zero as the wavenumbers grow, and the wavevectors here are "optimised without any constraints on their magnitudes or directions" (p. 21). A Fisher objective therefore has no minimiser on this class, and the finite, multiple scales in Fig. 4 can only come from the global-confusion term of Eq. 15. In other words, the bound sets the module scales. This argument belongs in the paper.

Critical weaknesses

No comparison code, no gridness score, one run. The claim that grid cells are "a highly efficient spatial code" needs a reference. The bound and the Monte Carlo error are reported only for optimised Fourier codes (Fig. 5), so no place-cell code, multi-field code or random Fourier code is ever evaluated under Eq. 15 at the same N = 73, C = 1 and σ_c = 0.08 (table, p. 24). Fig. 4 shows "6 selected neurons" (p. 8) without a selection rule, and no gridness score, shuffle threshold, phase or orientation distribution appears anywhere. By eye, about 8 of the 73 maps in Fig. 6 are clearly hexagonal and perhaps 7 more plausibly so, while the rest are stripes, single blobs or irregular multi-field patterns. Please provide gridness with a shuffle threshold for all neurons over several seeds, and the bound plus Monte Carlo error for a place-cell code, a multi-field code and the optimised code with randomised wavevector directions, all at matched N, C and σ_c.

The objective cannot see what a single neuron looks like and the hexagonal rate map, is then decided by the positivity penalty. That is the mechanism already known from Sorscher et al. (2019, not cited), Whittington et al. (2022) and Dorrell et al. (2022), whose Fourier class is the one used here

modules "separated by a factor of approximately 1.5, in agreement with the experimental data Hafting et al. (2005)" (p. 9), while Fig. 4 prints peaks at 0.61, 1.07 and 1.42 with ratios ×1.76 and ×1.34. Hafting et al. (2005) reported spacing increasing along the dorsoventral axis; the discrete modules with a ratio near 1.4 are Stensola et al. (2012)

The encoder receives the whole cue history and is corrupted once, "The trajectory history is encoded into the steadystate firing rates" (p. 3), so the code itself never integrates and never accumulates noise

Questions for authors
  1. At N = 73, C = 1 and σ_c = 0.08, what are the Ziv–Zakai bound and the Monte Carlo error of a place-cell code, a multi-field place code, and the optimised code with its wavevector directions randomised at fixed magnitudes and powers? This one comparison decides the efficiency claim.
  2. With the positivity term switched off, do the optimised wavevectors still form 60° triplets in discrete modules?
  3. What was the selection criterion in the hyperparameter sweep, and how many seeds produce the Fig. 4 structure?
  4. What is the final positivity violation of the Fig. 4 code, and what are P_c, the constant power and the mode powers?
  5. Where do the 81 encoders of Fig. 5 come from, and which decoder and sample size produced the Monte Carlo error?
  6. How is one length scale assigned per neuron in Fig. 4, and how many neurons fall in each of the three peaks?
3 comments
  • "The choice of the Ziv–Zakai bound is interesting, and more than the paper says" Thank you for your suggestion. We agree it fits well into the main text and thus we have added this under the paragraph "A well-defined optimisation problem". Further, as we explain in the text, the issue you highlight with the CRLB is not limited to just our particular parameterisation and can be linked to our "rope packing intuition".

  • W1: "Please provide gridness with a shuffle threshold for all neurons over several seeds, and the bound plus Monte Carlo error for a place-cell code, a multi-field code and the optimised code with randomised wavevector directions, all at matched N, C and sigma c." Thank you for these and your other comments. The concern that the computational results do not hold enough weight is shared by reviewer wreK, so we detail two new experiments in the general reply: a grid versus place cell comparison, as requested, and a non-parametric optimisation, as suggested by reviewer wreK. We have also included the gridness, orientation, phase and shuffled wavevector orientation analyses you suggested, alongside learnt tuning curves with different neuron numbers and sigma values, demonstrating grid-like tuning curves beyond the specific configuration in the text. We hope these help address your doubts. Following wreK’s restructuring comments and your other thoughtful suggestions, we have also clarified the intentions of each numerical experiment. We hope we have covered every point and apologise if we have missed any.

  • W2: "The objective cannot see what a single neuron looks like and the hexagonal rate map, is then decided by the positivity penalty." Thank you for highlighting this point. We agree that the objective does not uniquely determine individual neuron's tuning curves, and the original text did not make this clear. We have added a new paragraph to the main text called “A Function of the Neural Encoding Manifold Only” to make this explicit. In doing this, we have distinguished two types of transformations our objective cannot see: source measure preserving transformations and orthogonal transformations of neural activity. It is this second type which relates to the works you cited. We have now mentioned this in the text and added the Sorscher reference, alongside a reference to another of Dorrell’s papers, Convex Efficient Coding, which discusses more generally the uniqueness of individual neurons’ tuning curves and the effect of non-negativity. We agree that it was an oversight not to mention this and acknowledge the previous work.

  • W3: "modules "separated by a factor of approximately 1.5, in agreement with the experimental data Hafting et al. (2005)" (p. 9), while Fig. 4 prints peaks at 0.61, 1.07 and 1.42 with ratios ×1.76 and ×1.34" We apologise for the oversight on the citation and imprecise wording. We have corrected both issues accordingly.

W4: "The encoder receives the whole cue history and is corrupted once, "The trajectory history is encoded into the steadystate firing rates" (p. 3), so the code itself never integrates and never accumulates noise" Thank you for your observation. We see two possible interpretations of this point and address each below:

  • First, if you mean that noise in the self-motion cues does not accumulate, we apologise for any lack of clarity. The original text stated that ``at each time step t \in {1,\ldots,T} , self-motion velocity cues are corrupted by independent additive Gaussian noise'', as defined in Eq. 2. Noise is therefore added at every time step, not just once. This accumulation is reflected in the sensory drift term in Eq.~14: 2T\sigma_{\eta}^2 , rather than 2\sigma_{\eta}^2 . We agree that the original Figure 1 was ambiguous and have updated it accordingly. We have also changed our notation from y_t to \Delta y_t making it clear that the inputs are displacement updates.

  • Second, if you mean that the neural channel noise, n, does not accumulate, then we agree. This noise was intended to represent downstream transmission or readout noise, not recurrent integration. However, accumulating errors are present in our model through the noise added to each displacement update, as mentioned above. Although termed sensory noise, this can be interpreted as coming from several sources, including neural processing or recurrent integration. Importantly, this noise represents shared uncertainty of the represented position giving rise to information-limiting correlations (Moreno-Bote et al., 2014). We do not include noise independent to each neuron however the affects of this can be reduced through population averaging, and any remaining contribution that changes the best guess of the represented position could in principle also be incorporated into our sensory noise term. We acknowledge, however, that modeling these combined contributions as independent Gaussian noise is clearly not the most precise treatment and we of course do not explicitly model recurrent dynamics. Nevertheless, as described our model does include noise integration just in its simplest form. Furthermore, following reviewer wreK’s advice, we have substantially shortened this section reducing its emphasis, as it is not the paper’s main focus. We appreciate this comment, as it highlights a very subtle point, and agree that explicitly considering recurrent dynamics would be an important extension of this work.

Q1:"At N = 73, C = 1 and sigma c = 0.08, what are the Ziv–Zakai bound and the Monte Carlo error of a place-cell code, a multi-field place code, and the optimised code with its wavevector directions randomised at fixed magnitudes and powers?" As discussed, following your suggestion we have added an extra experiment addressing the direct grid vs multi/ single field place cell comparison.

Q2: "With the positivity term switched off, do the optimised wavevectors still form 60° triplets in discrete modules?" A preliminary test did not learn grid-like tuning curves, consistent with the previously mentioned literature. Q3: "What was the selection criterion in the hyperparameter sweep, and how many seeds produce the Fig. 4 structure?"} Fig. 4 was selected as due to it being the best performing encoding which met the constraints across the random seeds, with the same hyper parameters. Qualitatively, the other seeds produced similar tuning curves, with all seeds producing at least some strongly hexagonal tuning curves, like in Fig. 4, we will add however some seeds did have strong striped firing patterns in a small number of neurons. We selected the hyperparameters according to the papers efficient coding objective, i.e. picking the best learnt encoding which met the positivity threshold and minimised the coding loss. As mentioned, we have also added examples with different neuron numbers and noise levels to demonstrate robustness and to gvie more examples.

Q4: "What is the final positivity violation of the Fig. 4 code, and what are Pc, the constant power and the mode powers?" We selected the best-performing encoding that met the positivity violation threshold of e(-9) approx 0.000123. Its measured violation was 0.000119. The centred neural power, Pc, was 0.467, the DC power was 0.532, and the oscillatory power was 0.468. The small difference between centred and oscillatory power arises because the encoding is evaluated over a finite disc, where the cosine components may not have zero mean. As you can see the large DC power is clearly not ideal as it does not contribute to distinguishability. In the original text we mention in the discussion that the reduction to this parametric form should not be considered lossless, and this represents the clearest example as to why. Reviewer wreK picked up on this point and as requested we have now mentioned this earlier, to avoid any confusion.

Q5: "Where do the 81 encoders of Fig. 5 come from, and which decoder and sample size produced the Monte Carlo error?" We apologise for the number 81 and if it caused any confusion, we performed as many seeds as possible in the time. We have trained an additional 19 encoders such that we now have 100 points. The encoders are trained in exactly the same way as the encoder pictured in the paper, with different random seeds. The sample size was 200,000 points and the decoder a grid based maximum likelihood estimator.

Q6: "How is one length scale assigned per neuron in Fig. 4, and how many neurons fall in each of the three peaks?" Length scales were calculated from the autocorrelogram by identifying the six nearest peaks from the center that exceeded a threshold of 0.3. To avoid double-counting, these peaks were required to be at least 15 degrees apart; neurons failing to meet this criterion were excluded. We have added this detail to the appendix. In the paper’s figure, each short dashed line represents a neuron. There is some clustering, shown more clearly in the new length-scale versus orientation plot, but it is not exact, so assigning neurons to modules is approximate. We clearly do not see equal numbers in each module, with fewer as the scale increases and importantly the neurons in higher length scales tended to be less grid-like.

Comments on the paper

1 comment

First, we would like to really thank the reviewers for their careful reading of the manuscript and for their relevant and constructive comments that have been well appreciated. The major changes of the manuscript are twofold: (1) We significantly streamlined and clarified the main message of the manuscript and (2) we performed a lot of additional simulations (adding a non parametric optimisation and a direct grid vs place cell comparison) to strengthen the point of the paper, we hope these address all of your concerns.

  1. Streamlining the message:
  • we combined sections 2-4
  • we updated the main figure
  • we highlighted the Jensen's gaps
  • we added a simpler form of the ZZLB (which is more intuitive)
  • we computed the gridness score
  • we added a contributions section in the Discussion
  1. New Simulations:
  • In addition to the parametric encoding that we had in the first version, we added a non-parametric encoding which provides the full flexibility of encoding. We found that grid cells also emerge with this non-parametric encoding.
  • We also added an explicit comparison between grid-cells, multi-field place cells and place cells and obtained numerical evidence that grid-cells outperform multi-field place cells, which themselves outperform place cells
  • we performed additional analysis on the simulations (gridness, tuning curves at different noise levels, orientation and randomised wavevector orientation comparison) to make our claims more robust

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.