- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This work proposes a mathematical framework based on group theory, the group Fourier transform, and irreducible representations to understand why grid cells are a suitable representation for path integration. The authors formulate path integration as sequential composition of group elements, construct an RNN with interpretable computations and a provably asymptotically optimal number of hidden neurons under specific conditions, and then connect its neuronal activity patterns to grid cells, linking the mathematical description to a neural interpretation. The paper also offers predictions for 3D navigation and introduces irrep selectivity as a method for analyzing neural activity. I like this work overall. What I appreciate most is that the mathematical perspective is elegant and makes sense intuitively: on the one hand, we think of grid cells as supporting path integration; on the other, their periodicity and multiple spatial scales bring Fourier decomposition to mind. This work brings these two ideas, along with their relationship to group operations, into a single framework that can be analyzed mathematically. However, the paper really could be written more clearly, and the figures are also difficult to follow. Some symbols, definitions, and key steps are not explained well enough, so understanding the paper takes quite a bit of effort. It also needs to be clearer about which structures are chosen in advance, which results are derived, and what exactly “optimal” means. I would mainly view Section 5 as predictions about 3D representations, while Section 6 provides an interesting analysis tool and a preliminary application; their biological implications still need further validation. So, I would currently give the paper a 3. I appreciate the overall idea, and if these aspects are properly worked through, I would be willing to give it a 4.
Critical strengths
-
What I appreciate most is that the overall mathematical perspective is elegant and makes sense intuitively. Grid cells support path integration, while their periodicity and multiple spatial scales bring Fourier decomposition to mind. This work brings these two ideas, along with their relationship to group operations, into a single framework that can be analyzed mathematically, offering an insightful way to understand the problem.
-
The paper also turns this mathematical perspective into a concrete neural network construction. The network’s computations have an explicit interpretation, and its number of hidden neurons is asymptotically optimal under specific conditions. The authors then connect the activity patterns of these neurons to grid cells. I think this makes the link between the mathematical framework and neural representations more concrete, connecting the mathematical description to a neural interpretation.
Critical weaknesses
My biggest concern is that the hexagonal structure is already built into the model through the choice of a triangular lattice and sixfold rotational symmetry, so the resulting hexagonal tuning cannot serve as evidence explaining why grid cells have a hexagonal arrangement. What this shows is how to construct the corresponding representation given hexagonal symmetry; it does not answer why the navigation system would choose this symmetry. The paper also reports the emergence of multiple grid-cell modules, but this depends on how the model is divided into modules. The authors therefore need to explain this division more clearly, including which aspects are determined by the mathematical structure of the group and which depend on other choices made by the authors. For the latter, they need to explain why these choices would correspond to the actual neural system.
Questions for authors
I have questions about the following aspects:
-
How exactly are the two reference activity vectors determined, and why are they chosen this way? The allocentric neural representation seems to be constructed as localized bump activity, which the paper connects to place cells and head-direction cells, but there is no corresponding explanation for the egocentric representation. Since these choices affect the subsequent network construction and experiments, the main text needs to explain them and their rationale clearly. When the vectors are first introduced, it should also point to the specific appendix location containing their detailed definitions.
-
Fig. 2 needs a clearer explanation. The top and bottom rows look the same apart from their colors, yet the egocentric reference vector used in the actual experiments is not a Gaussian bump, while the figure makes it look like one. This distinction needs to be made explicit.
-
Some symbols and key steps are not explained sufficiently, for example in Eq. (5). Could the authors clarify what the symbols represent when they first appear and fill in the reasoning connecting group representations, the group Fourier transform, and the neural activity encoding?
-
Besides the illustrated position-selective, conjunctive, and heading-only units, what other activity types are present in the network? What proportion does each type account for, and are these distributions fully reported? The statistics described in the appendix cover only a subset with fixed labels. Please provide the existing counts and proportions for each category and explain how well this subset represents the full network.
2 comments
Thank you again very much for your thoughtful review and for your follow up.
Hexagonal Tuning
We are glad we could partially resolve your main concern with the normative arguments for using a triangular lattice. Again we want to emphasize that simply using a triangular lattice is not sufficient to see hexagonal tuning, a Fourier-basis is necessary. While we too still have questions, for example as we discuss future work should try to extend our framework from finite to continuous groups, we think we have provided a very clean and elegant framework that is a step in the right direction.
Multiple grid-cell modules
In our framework a grid-cell module is a group of neurons selective to the same irrep of . The group determines the available modules and their spatial structure. The allocentric code determines which are required for perfect path integration. For the reference code used in our experiments, modules at the ten largest distinct spacings and irreps with together account for of its power. We have made this clear in the main and added a new panel D to Figure 7.
Additionally, we should emphasize that if we allow for the RNN to accumulate error over long sequences (which is natural and observed biologically), then we can get away with using less irreps than needed for perfect reconstruction. Specifically, for the Gaussian bump allocentric code we use in our simulations, we find that restricting our construction to only the first few low-frequency irrep modules maintains very low error. See our new Fig. 15 in App. F. These low-frequency modules are also the modules with the largest and most distinct spacing. In this updated Appendix F section we also discuss how irrep modules relate to the lattice spacing. From an irrep we can determine a mathematical expression for the average spacing, defined as the distance between neighboring firing fields. We can then sort irreps from large spacing (low-frequency) to small spacing (high-frequency). Lastly we can then compute the ratio of spacing between consecutive modules, as done in Stensola et al. (2012). We find two really interesting results: (1) the ratio is independent of the discretization of the lattice in . (2) The average spacing ratio for the top four low-frequency modules (the most important ones for reducing path integration error) is with a standard deviation of 0.242. This is very close to the experimental estimate of from Stensola et al. (2012), as reanalyzed by Wei et al. (2015). While this is just a connection for now – we think this is worthy of future exploration.
Reference codes and original Fig. 2.
We choose the allocentric reference as a localized spatial bump with nonuniform heading amplitudes, motivated by position- and heading-selective population activity. The egocentric reference is a generic random code, sampled to have invertible Fourier coefficients on the retained irreps. Its particular tuning shape is a modeling choice, without a claimed biological correspondence. Appendix F gives the definitions and sampling procedure, and we added a visualization in Fig.14. We would like to emphasize that while the egocentric reference will affect the weights, it actually doesn’t affect the tuning curves because the weights effectively remove it as seen in equation 7 by .
Fig. 2 illustrates the regular group action schematically: the egocentric bump shown there is not the random reference code used in the experiments. To avoid confusion we have removed it leaving just the allocentric reference code.
Notation
A full introduction to group theory and representation theory in the main would take too much space. Around Eq. (5) we now point to Appendix A “Groups and Their Fourier Transforms” which serves as a compact primer to these mathematical topics. We have added citations here to textbooks with formal introductions.
Activity types and sampling
Our classification distinguishes position-only, heading-only, conjunctive, and units unmodulated in both marginals. The counts and proportions are now shown in Fig. 18.
Thank you again for your review and for your follow up. We hope in light of these additional experiments and explanations you will raise your score and continue to advocate for the acceptance of our paper.
Thank you for your response. While I appreciate the clarification, my main concern is not fully resolved. Even if a triangular lattice is not sufficient by itself to produce hexagonal tuning, it already imposes a geometric structure closely related to the observed pattern. It therefore remains difficult to argue that hexagonal tuning emerges from path integration independently of the model’s assumptions.
My other questions have not yet been addressed. As most of my concerns remain unresolved, I will maintain my current rating for now.