Pith. sign in

REVIEW 5 major objections 5 minor 13 references

The Hippocampal Place Field Gradient: An Eigenmode Theory Linking Grid Cell Projections to Multiscale Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a frequency-dependent decay in grid-to-place weights produces the continuous place field gradient, and that the optimal place field size for few-shot learning is proportional to the task's characteristic spatial…

desk verdict A genuinely interesting mechanistic proposal for the place-field gradient, but the main quantitative result rests on an assumed eigenvalue spectrum rather than a derivation from the model's own weights. read the letter →

arxiv 2506.04943 v1 pith:23QVZD3X submitted 2025-06-05 q-bio.NC

Shujun Zhou , Guozhang Chen This is my paper
classification q-bio.NC
keywords placefieldgradientgridcellshippocampusmedialentorhinalcortexeigenmodefew-shotlearninginductivebiaskernelregression
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to explain the continuous dorsal-ventral gradient of hippocampal place field sizes from the discretely scaled grid cell modules of the medial entorhinal cortex, and to show that this gradient is functionally meaningful for learning. The proposed mechanism is a linear readout: place cells are linear projections of grid cell activity, the grid basis is the Fourier eigenbasis of any translation-invariant spatial code, and the amplitude of each grid-to-place weight decays exponentially with spatial frequency at a rate set by the place field width. That single decay rate turns discrete modules into a continuous spectrum of field sizes and also sets the inductive bias of the population code. Using kernel regression theory, the paper derives an optimal place field width for few-shot learning, $\sigma_p^{\mathrm{opt}} \propto D$, where $D$ is the characteristic region size of the task, which would make the gradient a multiscale code matched to environment structure. If correct, this yields testable anatomical predictions for dorsal versus ventral MEC projections and suggests a design principle for artificial few-shot learners.

What carries the argument

The engine of the argument is the covariance kernel $\Sigma(x,x') = \mathbf{p}(x)^\top \mathbf{p}(x')$ of the place-cell population, together with its eigen-decomposition. Translational invariance on a periodic track forces the kernel to commute with the translation operator, so its eigenmodes are Fourier modes; the grid cells are identified with those modes, which diagonalizes the readout problem. The closed-form projection of a Gaussian place field onto a Fourier mode yields connection amplitudes decaying as $\exp[-\sigma_p^2 (2\pi k/L)^2/2]$, and the same $\sigma_p$ appears in the kernel eigenvalues $\lambda_k = \exp[-\sigma_p^2 (k 2\pi/L)^2]$ used in the generalization-error calculation. The final optimization—maximizing the single-mode denominator in the few-shot error—gives $\sigma_p^{\mathrm{opt}} = L/(2\sqrt{2\pi}\, n) \propto D$.

What would settle it

Compute $W^\top W$ using the deterministic least-squares weights of Section 2.2; if the kernel eigenvalues are not the assumed diagonal exponential form, then the analytic optimum changes. Experimentally, a high-resolution MEC-to-CA1/CA3 projectome would falsify the anatomical prediction if dorsal CA1/CA3 do not receive mixed dorsal and ventral MEC input while ventral CA1/CA3 are dominated by ventral MEC.

Watch

Extended reading notes

Core claim

The central discovery is that the same object—the spectrum of the place-cell covariance kernel—explains both the anatomy of grid-to-place projections and the learning efficiency of the place code. On a periodic track of length $L$, any translationally invariant population covariance commutes with the translation operator, so its eigenmodes are Fourier modes; since grid cells fire in periodic patterns, they are exactly these eigenmodes. For Gaussian place fields of width $\sigma_p$, the optimal linear projection from grid module $k$ to place cell $i$ has amplitude $\frac{A}{L}\exp[-\sigma_p^2 (2\pi k/L)^2/2]$, so the field width acts as a frequency-dependent decay on connection weights and a discrete set of grid modules produces a continuous range of field sizes. Substituting the resulting kernel eigenvalues $\lambda_k = \exp[-\sigma_p^2 k^2 (2\pi/L)^2]$ into ridge-regression generalization theory, and keeping the task mode $n \approx L/(2D)$ that dominates a region-based task, gives $\sigma_p^{\mathrm{opt}} = L/(2\sqrt{2\pi}\, n) \propto D$. The paper concludes that the dorsal-ventral gradient is a multiscale inductive-bias code: small fields give precision when data are abundant, large fields give few-shot generalization, and the best scale is set by the spatial structure of the task.

Load-bearing premise

The load-bearing premise is Appendix B's assertion that grid-to-place weights are independent across place cells with covariance $C^2 \delta_{ij} \exp[-\sigma_p^2 k^2 (2\pi/L)^2]$, which makes the population kernel diagonal; this is assumed rather than derived from the deterministic least-squares weights given earlier, and if the real population covariance is not diagonal the analytic result $\sigma_p \propto D$ is not established.

Editorial extensions

If this is right

  • A discrete set of grid modules is sufficient: the continuous place field gradient can be produced entirely by the exponential decay rate of connection weights, with no need for continuously graded grid scales.
  • The optimal place field width for few-shot learning is set by the task's dominant spatial scale $D$, and this optimum is approximately independent of the number of training samples in the few-shot regime.
  • The anatomical prediction follows: dorsal CA1/CA3 place cells should pool both dorsal and ventral MEC inputs, while ventral CA1/CA3 should be dominated by ventral MEC, testable with projectome data.
  • When the environment is larger than the largest grid module, place cells should acquire multiple firing fields through recruitment of higher-order Fourier components.
  • For artificial systems, the theory implies that fixed positional encodings can be given different effective scales by modulating the weights from the encoding layer, shaping few-shot generalization without changing the encoding itself.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • [Editorial inference] If the spectral-alignment principle is general, then in non-spatial “cognitive spaces” the optimal tuning width of a population code should track the dominant scale of the relational structure, so the same eigenmode logic would predict analogous gradients in abstract representations.
  • [Editorial inference] The single-mode approximation used for the analytic optimum is a strong filter; for tasks with broad spectra, the optimal code may need a mixture of scales, and a multiscale population should robustly outperform any one-width population across tasks—a direct, testable extension of the paper's code-task alignment logic.
  • [Editorial inference] The independence assumption on grid-to-place weights in Appendix B could be tested by measuring or simulating the true population covariance from the fitted projection weights; if the covariance is not diagonal, the analytic scaling may still hold in simulations, but it would not follow from the derivation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes a theory of the hippocampal dorsoventral place-field gradient. It models place fields as linear projections of grid-cell Fourier modes, shows that least-squares weights for Gaussian fields decay exponentially with spatial frequency, and interprets this as a mechanism for generating a continuum of field sizes from discrete grid modules. It then uses kernel-regression generalization theory to argue that field width controls a precision–generalization trade-off, derives an optimal width σ_p^opt = L/(2√(2π)n) ∝ D for few-shot learning, and supports this with simulations (Figures 3–5). The paper also makes anatomical predictions about dorsal/ventral MEC projections and discusses implications for machine learning.

Significance. The qualitative center of the paper—small fields optimize precision when data are abundant, large fields generalize better when data are scarce, and the best field size grows with the spatial scale of the task—is clearly illustrated by the simulations and is a plausible functional rationale for the dorsoventral gradient. The paper also offers concrete, testable anatomical predictions (Section 2.3) and connects hippocampal coding to kernel-regression theory, which gives the framework broader currency. However, the quantitative claims rest on an eigenvalue spectrum that is assumed rather than derived (Appendix B) and on a derivation whose prefactor is inconsistent with the full maximization (Appendix D), so the analytic results should be treated as provisional. If the spectrum question is fixed, the linear σ_p^opt–D scaling could be a useful contribution.

major comments (5)
  1. [Appendix B, Eq. (41)] The eigenvalue spectrum λ_k = exp[-σ_p² k²(2π/L)²] is assumed, not derived. With the least-squares weights in Eqs. (9)–(10), WᵀW = Σ_i w_i w_iᵀ is a sum of rank-one deterministic functions of the place-field centers μ_i; for finite n_p its off-diagonal blocks do not vanish and its eigenvectors are not exactly Fourier modes. Eq. (41) replaces this object with an independent diagonal covariance, which is an additional assumption, not a consequence of the model. Because κ, γ, and the minimization leading to Eq. (24) all depend on this spectrum, the analytic derivation is at best an infinite-population result. The paper should derive the spectrum from Eqs. (9)–(10) or from the direct population covariance with random μ_i and quantify finite-n_p corrections, or state that Eq. (24) is conditional on the assumed spectrum.
  2. [Section 2.2, Eqs. (9)–(10)] The claim that a frequency-dependent decay of connection weights determines place field size is circular. The weights are computed by least-squares projection of a pre-assigned Gaussian place field onto the Fourier basis, so exp[-σ_p²(2πk/L)²/2] is simply the Fourier transform of the assumed Gaussian. The construction runs from field size to weights, not from weights to field size. To support the explanatory arrow, the model needs an independent mechanism (for example, a plasticity rule or a connectivity constraint) that sets the weight spectrum and then produces the field, rather than reading Eq. (9) backwards.
  3. [Section 3.5, Eqs. (20)–(24)] The algebraic derivation contains a constant error and an inconsistent treatment of the (1−γ) factor. Evaluating Eq. (20) with ∫₀^∞ e^{-u²}du = √π/2 gives κ = L/(2√π σ_p), not Eq. (21)'s L/(√(2π)σ_p); this factor of √2 propagates into the constant C in Eq. (23). In addition, Eqs. (22)–(23) minimize 1/(Cσ_p e^{−...}+1)² by dropping the (1−γ) prefactor, whereas Appendix D retains it and obtains σ_p^opt = L/(2√3π n), which differs numerically from Eq. (24). Since the prefactor in Eq. (24) is part of the paper's specific prediction, the two derivations must be reconciled.
  4. [Section 3.4, Eqs. (15)–(17), Appendix A] The definition of γ is inconsistent with the derivation. The reproduced Bordelon–Pehlevan/Canatar derivation requires γ_eff = S Σ_k λ_k²/(λ_k S+κ)², while Eq. (17) omits the factor S. Appendix A explicitly notes this mismatch. Because γ multiplies the whole error expression in Eq. (15), the formula as printed is not a correct rendering of the cited framework. Please correct the definition and re-derive the few-shot expressions consistently.
  5. [Appendix D, Eqs. (50)–(63)] This appendix is not in publishable form. It contains phrases such as "the prompt text uses..." and "following the prompt's derivation steps," which are artifacts of an analysis prompt rather than language appropriate for a paper. Independently of the style, the appendix's final constant 1/(2√3πn) contradicts Eq. (24). The appendix should be rewritten as a self-contained derivation with consistent constants and no meta-commentary.
minor comments (5)
  1. [Throughout] There are typographical errors: "activtiy" (Section 2.1), "spectual" (Section 3.5), "projecitions" (Section 2.3), and "the the" (Section 1).
  2. [Section 3.5, after Eq. (24)] The phrase "the three factors affecting the learning performance i.e. S, λk, µk" should read "νk" for the task coefficients.
  3. [Appendix C] Appendix C refers to Figure 6, but no Figure 6 appears in the manuscript text; include the figure or remove the reference.
  4. [Figures 3–5] Please specify the simulation hyperparameters (numbers of grid modules and place cells, training sample sizes S, learning rates, number of runs, and error bars) to make the empirical claims reproducible.
  5. [Section 3.3] The sentence "Although the place field size σp can be set arbitrarily. it is limited..." needs punctuation correction: "arbitrarily, it is limited...".

Circularity Check

2 steps flagged · score 6.0 of 10

The analytic optimal-width derivation assumes the very σ_p-dependent spectrum it then predicts; Sec. 2.2's weight-decay relation is the Fourier identity read backwards.

  1. self definitional [Sec. 2.2, Eqs. (9)-(10) and following paragraph]
    "Thus, the amplitude of the projection from the grid module with frequency k to place cell i (aik = q W 2 i,2k−1 + W 2 i,2k) decays exponentially with spatial frequency, with the place field size σp controlling the rate of the decay: aik ∝ exp[− σ2 p(2πk/L)2 2 ]. This has a clear theoretical implication: despite the fact that grid codes themselves operate on discrete spatial scales, continuous tuning of place field size can be achieved by systematically adjusting the decay of weights across spatial frequencies."

    Equations 9-10 compute the weight matrix W by projecting a Gaussian place field of width σp onto Fourier grid modes; σp is the input, and W is the output. The paper then reads the resulting exponential decay of W as 'achieving' a place field of size σp, reversing the direction of the derivation. Because the readout is the least-squares projection, the amplitude relation is exactly the Fourier transform of the assumed Gaussian tuning curve; no independent information about how weights generate field sizes is added.

  2. self definitional [Appendix B, Eqs. (41)-(42), used in Sec. 3.5, Eqs. (18)-(24)]
    "To generate the desired Gaussian place fields, and assuming that the connection weights between grid cells and place cells are independent across place cells, we have, the weight matrix W must satisfy: E[Wi,2k−1Wj,2k−1] = C 2 δij exp[−σ2 pk2(2π/L)2] ... Therefore, the matrix W ⊤W takes the diagonal form ... By setting C = 1, we directly recover Equation 18."

    Equation 41 postulates a diagonal W^T W with eigenvalues exp[−σ_p^2 k^2(2π/L)^2], and Equation 42 then reads off λ_k; Equation 18 is true by construction. This is not derived from the paper's own deterministic least-squares weights, Eqs. 9-10, for which W^T W is non-diagonal and depends on the place-field centers μ_i. The subsequent derivation of κ (Eq. 21) and of σ_p^opt ∝ D (Eq. 24) uses only this assumed spectrum, so the analytic prediction is forced by the σ_p-dependence inserted into the assumption rather than established by the model's stated weight structure.

full rationale

The paper contains no load-bearing self-citation: the theoretical framework it relies on (Bordelon and Pehlevan 2022; Canatar et al. 2021; Sorscher et al. 2023) is independent prior work. The Sec. 2.2 relation between weight decay and place-field width is a genuine mathematical identity for Gaussian fields under least-squares readout; presenting it as 'weight decay determines field size' is a modeling restatement rather than a data-driven prediction. The substantive circularity is in the analytic optimal-width derivation. Appendix B obtains λ_k = exp[−σ_p^2 k^2(2π/L)^2] by postulating a diagonal W^T W with exactly that exponential σ_p dependence (Eq. 41), rather than by computing the spectrum of the deterministic least-squares weights of Eqs. 9-10. Thus Eq. 18 is true by assumption, and Eqs. 21-24 propagate that assumed spectrum into κ and σ_p^opt ∝ D. The analytic 'prediction' that optimal field width scales with region size is therefore contingent on an ansatz that already contains the σ_p-dependence the theory claims to derive. The simulations in Figures 4-5, which use the full task structure and finite place-cell populations, provide independent support for the qualitative scaling, so the circularity is partial rather than total.

Assumptions & free parameters 3 free parameters · 9 assumptions · 0 invented entities

The ledger shows the framework pulls most of its content from prior literature: translation invariance and Fourier eigenmodes (Sorscher et al. 2023), the kernel regression error equations (Canatar et al. 2021, Bordelon and Pehlevan 2022), and the Gaussian place field assumption. The paper's genuinely new claim, that optimal place field width scales with task scale, depends mainly on three inputs: the Gaussian shape assumption, the Appendix B weight-covariance assumption that manufactures the eigenvalue spectrum, and the single-mode task approximation. No invented entities are introduced. The free parameters are modest (a normalization C set to 1, a module cutoff k_max = 8, and unspecified training hyperparameters), but the assumed rather than derived eigenvalue spectrum is the largest unearned input.

free parameters (3)
  • C (grid-to-place weight covariance scale) = set to 1
    Appendix B, Eq 41 normalizes the assumed weight covariance to C^2 exp[-sigma_p^2 k^2 (2pi/L)^2] and then fixes C = 1 to recover Eq 18. C cancels in the optimal-width ratio, but the exponential form itself is an assumption that loads the result.
  • k_max (highest grid module included) = 8 (n_g/2 = 8)
    Sec 3.3 truncates the grid module set at the 8th harmonic in the simulations. The authors state that for sigma_p much smaller than L/k_max the modules cannot support the place field, so the claimed continuous gradient is resolution-limited by this chosen cutoff.
  • n_p (place cell count) and training protocol = not stated
    The population size and the training details for Figures 3-5 (architecture, optimizer, learning rate, epochs, trials) are never given, and results are presented without error bars, so the simulation outcomes are under-specified.
assumptions (9)
  • domain assumption Spatial covariance of the population code is translationally invariant (Eq 2)
    Invoked in Sec 2.1 to conclude the covariance kernel commutes with the translation operator, yielding Fourier eigenmodes. Drawn from Sorscher et al. 2023 but assumed without physiological justification for the hippocampal code.
  • domain assumption Periodic boundary conditions on a 1D circular track
    Sec 2.1 imposes f(x) = f(x + L). The paper acknowledges in its limitations that environmental boundaries, which strongly shape real place cells, are ignored.
  • domain assumption Place fields are Gaussian tuning curves with width sigma_p
    Sec 2.2 states p_i(x) proportional to exp[-(x - mu_i)^2 / (2 sigma_p^2)]. All downstream results (weight decay, eigenvalues, optimal width) inherit this shape assumption.
  • domain assumption Grid cells are pure Fourier modes cos(2pi kx/L), sin(2pi kx/L) in 1D
    Sec 2.2 defines grid tuning as Fourier modes. Real MEC grid cells are 2D hexagonal lattices, and the claimed extension to d dimensions is asserted without proof in Sec 2.1.
  • domain assumption The grid-to-place mapping is the least-squares linear readout (Eq 8)
    Sec 2.2 minimizes the squared error between Gaussian place fields and the linear readout W g(x). This optimality assumption is what converts Gaussian fields into exponential weight decay; biological plasticity is mentioned only as future work.
  • standard math Spectral bias / kernel regression framework of Canatar et al. and Bordelon-Pehlevan
    Sec 3.4 and Appendix A import the self-consistent equations for kappa, gamma, and E_g from Canatar et al. 2021 and Bordelon and Pehlevan 2022, including the replica/DMFT averaging over training sets.
  • ad hoc to paper Weight independence across place cells with assumed exponential covariance (Eq 41)
    Appendix B postulates E[W_{i,2k-1}W_{j,2k-1}] = C^2 delta_ij exp[-sigma_p^2 k^2 (2pi/L)^2], which directly yields the eigenvalue spectrum lambda_k used in the optimal-width derivation. This contradicts the deterministic weights of Eqs 9-10, which have cos/sin(2pi k mu_i/L) structure and would generate off-diagonal correlations.
  • ad hoc to paper Single-mode approximation of the task spectrum (nu_k = nu_n delta_{k,n})
    Sec 3.5 and Appendix C collapse the task onto its dominant Fourier mode n approximately L/(2D). Acknowledged as a simplification, with Figure 6 and Figure 5 offered as numerical support.
  • domain assumption Small-sample approximation S much less than kappa, with alpha = 0
    Sec 3.5 derives kappa approximately L/(sqrt(2pi) sigma_p) only in the few-shot regime with no regularization. The claim that sigma_p^opt is independent of S is a consequence of this limit, not of the full self-consistent equations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Hippocampal Place Field Gradient: An Eigenmode Theory Linking Grid Cell Projections to Multiscale Learning." pith.science (2026). https://pith.science/paper/23QVZD3X

@misc{pith2026250604943,
  author       = {Pith},
  title        = {Pith review of: The Hippocampal Place Field Gradient: An Eigenmode Theory Linking Grid Cell Projections to Multiscale Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23QVZD3X}},
  note         = {Machine review of arXiv:2506.04943}
}
read the original abstract

The hippocampus encodes space through a striking gradient of place field sizes along its dorsal-ventral axis, yet the principles generating this continuous gradient from discrete grid cell inputs remain debated. We propose a unified theoretical framework establishing that hippocampal place fields arise naturally as linear projections of grid cell population activity, interpretable as eigenmodes. Critically, we demonstrate that a frequency-dependent decay of these grid-to-place connection weights naturally transforms inputs from discrete grid modules into a continuous spectrum of place field sizes. This multiscale organization is functionally significant: we reveal it shapes the inductive bias of the population code, balancing a fundamental trade-off between precision and generalization. Mathematical analysis and simulations demonstrate an optimal place field size for few-shot learning, which scales with environment structure. Our results offer a principled explanation for the place field gradient and generate testable predictions, bridging anatomical connectivity with adaptive learning in both biological and artificial intelligence.

Figures

Figures reproduced from arXiv: 2506.04943 by the authors.

Figure 1
Figure 1. (a) Grid-to-place cell weights as a function of place field center [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Model architecture and context-dependent task design. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Test accuracy under varying training sample sizes and place field widths. The length of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (a) Relationship between test accuracy and place field width. (b) Optimal place field width [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Performance landscape under varying place field sizes and region widths. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: illustrates the squared projection coefficients ν 2 k for a representative instance of the task structure used in our simulations. The total length of the environment is L = 100 and the region width is D = 25. In this case, the task function y(x) changes its value ever…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 7 canonical work pages

  1. [3]

    Abdulkadir Canatar, Blake Bordelon, and Cengiz Pehlevan

    doi: 10.1038/s41593-017-0055-3. Abdulkadir Canatar, Blake Bordelon, and Cengiz Pehlevan. Spectral bias and task-model align- ment explain generalization in kernel regression and infinitely wide neural networks. Nature Communications, 12(1):2914, May

    doi:10.1038/s41593-017-0055-3

  2. [5]

    doi: 10.1371/journal.pcbi.1008835

    ISSN 1553-7358. doi: 10.1371/journal.pcbi.1008835. Torkel Hafting, Marianne Fyhn, Sturla Molden, May-Britt Moser, and Edvard I Moser. Microstructure of a spatial map in the entorhinal cortex. 436,

    work page doi:10.1371/journal.pcbi.1008835

  3. [8]

    Thus, the factor (1 − γ) in Equation (6) of the main paper corresponds to (1 − SγEq.8) in the context of this derivation

    Comparing the derived γeff with γ from Equation (8) of the main paper, we see γeff = S · γEq.8. Thus, the factor (1 − γ) in Equation (6) of the main paper corresponds to (1 − SγEq.8) in the context of this derivation. This implies that either γ in Equation (6) is implicitly defined as S P k λ2 k (λkS+κ)2 , or Equation (8) in the main paper defines a quant...

    work page 2021

  4. [10]

    URL https://www.annualreviews.org/content/ journals/10.1146/annurev-neuro-102423-100258

    1146/annurev-neuro-102423-100258. URL https://www.annualreviews.org/content/ journals/10.1146/annurev-neuro-102423-100258 . Shou Qiu, Yachuang Hu, Yiming Huang, Taosha Gao, Xiaofei Wang, Danying Wang, Biyu Ren, Xiaoxue Shi, Yu Chen, Xinran Wang, Dan Wang, Luyao Han, Yikai Liang, Dechen Liu, Qingxu Liu, Li Deng, Zhaoqin Chen, Lijie Zhan, Tianzhi Chen, Yuzh...

    work page doi:10.1146/annurev-neuro-102423-100258

  5. [1993]

    A Derivation of the Generalization Error The derivation of the average generalization error for kernel regression presented here follows the framework established by Bordelon and Pehlevan [2022], which builds upon work by Canatar et al. [2021]. We adapt the notation to match the main paper. The average generalization error Eg is defined as the expected sq...

    work page 2022

  6. [2008]

    doi: 10.1126/science.1157086

    ISSN 0036-8075, 1095-9203. doi: 10.1126/science.1157086. Boyang Li, Yulin Wu, and Nuoxian Huang. GridPE: Unifying Positional Encoding in Transformers with a Grid Cell-Inspired Framework, June

    work page doi:10.1126/science.1157086

  7. [2012]

    doi: 10.1038/nature11649

    ISSN 0028-0836, 1476-4687. doi: 10.1038/nature11649. Bryan A. Strange, Menno P. Witter, Ed S. Lein, and Edvard I. Moser. Functional organization of the hippocampal longitudinal axis. Nature Reviews Neuroscience, 15(10):655–669, October

    doi:10.1038/nature11649

  8. [2014]

    doi: 10.1038/nrn3785

    ISSN 1471-003X, 1471-0048. doi: 10.1038/nrn3785. James C. R. Whittington, Joseph Warren, and Timothy E. J. Behrens. Relating transformers to models and neural representations of the hippocampal formation, March

    work page doi:10.1038/nrn3785

Show all 13 references
  1. [2018]

    doi: 10.1016/j

    ISSN 08966273. doi: 10.1016/j. neuron.2018.10.002. Blake Bordelon and Cengiz Pehlevan. Population codes enable learning from few examples by shaping inductive bias. eLife, 11:e78606, December

    2018 doi

  2. [2021]

    doi: 10.1038/s41467-021-23103-1

    ISSN 2041-1723. doi: 10.1038/s41467-021-23103-1. Daniela Gandolfi, Jonathan Mapelli, Sergio M. G. Solinas, Paul Triebkorn, Egidio D’Angelo, Viktor Jirsa, and Michele Migliore. Full-scale scaffold model of the human hippocampus CA1 area. Nature Computational Science, 3(3):264–2...

    doi

  3. [2022]

    doi: 10.7554/eLife.78606

    ISSN 2050-084X. doi: 10.7554/eLife.78606. Caitlin S. Mallory, Caitlin S. Mallory, Kiah Hardcastle, Kiah Hardcastle, Jason S. Bant, Jason S. Bant, Lisa M. Giocomo, and Lisa M. Giocomo. Grid scale drives the scale and long-term stability of place maps. Nature Neuroscience, 21(2)...

    doi

  4. [2023]

    doi: 10.1016/j.neuron.2022.10.003

    ISSN 0896-6273. doi: 10.1016/j.neuron.2022.10.003. Hanne Stensola, Tor Stensola, Trygve Solstad, Kristian Frøland, May-Britt Moser, and Edvard I. Moser. The entorhinal grid map is discretized. Nature, 492(7427):72–78, December

    2022 doi

  5. [2024]

    Ben Sorscher, Gabriel C

    doi: 10.1073/pnas.2312281120. Ben Sorscher, Gabriel C. Mel, Samuel A. Ocko, Lisa M. Giocomo, and Surya Ganguli. A unified theory for the computational and mechanistic origins of grid cells. Neuron, 111(1):121–137.e13, January

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.