Coherently remapping toroidal cells but not Grid cells are responsible for path integration in virtual agents

Schøyen, Vemund; Pettersen, Markus Borud; Holzhausen, Konstantin; Fyhn, Marianne; Malthe-Sørenssen, Anders; Lepperød, Mikkel Elle

doi:10.1016/j.isci.2023.108102

Cited by 7 publications

(7 citation statements)

References 57 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Pruning velocity inputs further allowed us to disentangle the contributions of different cell types to path integration. By comparing the representations of a pruned network to a non-pruned network, we could directly ascertain that path integration was not critically dependent on grid cells, but seemingly rather on band-type units, echoing other recent findings [Schøyen et al, 2023]. The inverse correlation between the contribution to path integration and the mean sample grid score further reinforces the notion that grid cells might not be as crucial for path integration as previously thought.…”

Section: Resultssupporting

confidence: 82%

“…Recently, it has also been shown that grid-like representations emerge in neural networks trained to path integrate [Cueva and Wei, 2018, Banino et al, 2018, Sorscher et al, 2022, Whittington et al, 2020, Xu et al, 2022, Dorrell et al, 2022, Schaeffer et al, 2023, which has been taken as evidence for grid cells performing path integration. However, this argument is based on correlation; under interventional cell ablations grid cells are as important for path integration as randomly selected cells [Nayebi et al, 2021] while band cells are significantly more important [Schøyen et al, 2023]. Moreover, these models are typically complex, featuring different architectures and activation functions, interacting label cell types (e.g.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Self-Supervised Grid Cells Without Path Integration

Pettersen,

Schøyen,

Østby

et al. 2024

Preprint

Self Cite

View full text Add to dashboard Cite

Grid cells, found in the medial Entorhinal Cortex, are known for their regular spatial firing patterns. These cells have been proposed as the neural solution to a range of computational tasks, from performing path integration, to serving as a metric for space. Their exact function, however, remains fiercely debated. In this work, we explore the consequences of demanding distance preservation over small spatial scales in networks subject to a capacity constraint. We consider two distinct self-supervised models, a feedforward network that learns to solve a purely spatial encoding task, and a recurrent network that solves the same problem during path integration. Surprisingly, we find that this task leads to the emergence of highly grid cell-like representations in both networks. However, the recurrent network also features units with band-like representations. We subsequently prune velocity inputs to subsets of recurrent units, and find that their grid score is negatively correlated with path integration contribution. Thus, grid cells emerge without path integration in the feedforward network, and they appear substantially less important than band cells for path integration in the recurrent network. Our work provides a minimal model for learning grid-like spatial representations, and questions the role of grid cells as neural path integrators. Instead, it seems that distance preservation and high population capacity is a more likely candidate task for learning grid cells in artificial neural networks.

show abstract

Section: Resultssupporting

confidence: 82%

Section: Introductionmentioning

confidence: 99%

Self-Supervised Grid Cells Without Path Integration

Pettersen,

Schøyen,

Østby

et al. 2024

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…4d) and e)), further hinting at different functional roles. That band cells contribute to path integration, has also been found in other neural network models [Schøyen et al, 2023]. As with the feedforward model, recurrent responses accurately capture the desired similarity structure (Fig.…”

Section: Recurrent Network Learn Place-and Band-like Representationssupporting

confidence: 69%

Learning Conjunctive Representations

Pettersen,

Rogge,

Lepperød

2024

Preprint

Self Cite

View full text Add to dashboard Cite

Hippocampal place cells are known for their spatially selective firing patterns, which has led to the suggestion that they encode an animal’s location. However, place cells also respond to contextual cues, such as smell. Furthermore, they have the ability to remap, wherein the firing fields and rates of cells change in response to environmental changes. How place cell responses emerge, and how these representations remap is not fully understood. In this work, we propose a similarity-based objective function that translates proximity in space, to proximity in representation. We show that a neural network trained to minimize the proposed objective learns place-like representations. We also show that the proposed objective is trivially extended to include other sources of information, such as context information, in the same way. When trained to encode multiple contexts, networks learn distinct representations, exhibiting remapping behaviors between contexts. The proposed objective is invariant to orthogonal transformations. Such transformations of the original trained representation (e.g. rotations), therefore yield new representations distinct from the original, without explicit relearning, akin to remapping. Our findings shed new light on the formation and encoding properties of place cells, and also demonstrate an interesting case of representational reuse.

show abstract

“…As our model only learns border-type recurrent representations, our findings raise question concerning the necessity of grid cells for path integration, as well as the causal relationship between place cells and grid cells. That grid cells may not be required to do path integration has also been shown in other recent normative models [36].…”

Section: Discussionmentioning

confidence: 59%

“…To the best of our knowledge, however, no normative models have seriously tackled the problem of place cell formation and remapping. Only some address remapping, but do so for other cell types [22], [35], [36].…”

Section: Introductionmentioning

confidence: 99%

Decoding the Cognitive map: Learning place cells and remapping

Pettersen,

Schøyen,

Malthe-Sørenssen

et al. 2024

Preprint

Self Cite

View full text Add to dashboard Cite

Hippocampal place cells are known for their spatially selective firing and are believed to encode an animal’s location while forming part of a cognitive map of space. These cells exhibit marked tuning curve and rate changes when an animal’s environment is sufficiently manipulated, in a process known as remapping. Place cells are accompanied by many other spatially tuned cells such as border cells and grid cells, but how these cells interact during navigation and remapping is unknown. In this work, we build a normative place cell model wherein a neural network is tasked with accurate position reconstruction and path integration. Motivated by the notion of a cognitive map, the network’s position is estimated directly from its learned representations. To obtain a position estimate, we propose a non-trainable decoding scheme applied to network output units, inspired by the localized firing patterns of place cells. We find that output units learn place-like spatial representations, while upstream recurrent units become boundary-tuned. When the network is trained to perform the same task in multiple simulated environments, its place-like units learn to remap like biological place cells, displaying global, geometric and rate remapping. These remapping abilities appear to be supported by rate changes in upstream units. While the model does not learn grid-like units, its place cell centers form clusters organized in a hexagonal lattice in open fields. This suggests a potential mechanism for the interaction between place cells, border cells, and grid cells. Our model provides a normative framework for learning spatial representations previously reserved for biological place cells, providing new insight into place cell field formation and remapping.

show abstract

Coherently remapping toroidal cells but not Grid cells are responsible for path integration in virtual agents

Cited by 7 publications

References 57 publications

Self-Supervised Grid Cells Without Path Integration

Self-Supervised Grid Cells Without Path Integration

Learning Conjunctive Representations

Decoding the Cognitive map: Learning place cells and remapping

Contact Info

Product

Resources

About