Coherently Remapping Toroidal Cells but Not Grid Cells Are Responsible for Path Integration in Virtual Agents

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

doi:10.2139/ssrn.4309105

Cited by 2 publications

(7 citation statements)

References 40 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…S6B). These results are consistent with similar analysis that was performed previously on single agent RNNs [31, 32]. When performing the same analysis on the population activations of dual agent RNNs, we find differences in the persistent homology.…”

Section: Resultssupporting

confidence: 92%

“…To understand their importance in dual agent path integration, we perform targeted ablations, removing units with the highest grid, border, and band scores (Appendix C.3). We compare the decoding error of these RNNs to RNNs with randomly removed units, as this has been found to be a strong baseline [31, 38]. Unlike single agent RNNs, which have a significant increase in decoding error relative to the random baseline with the ablation of all functional classes, dual agent RNNs are less sensitive to the ablation of any one functional class (Fig.…”

Section: Resultsmentioning

confidence: 99%

“…The construction of rate maps, which visualize activity of individual neurons, as a function of physical location, is a standard tool used to understand the neural representations underlying spatial navigation. In the context of RNNs trained on single agent path integration, this tool has been used to discover grid, border, and band cells [27–29, 31, 32]. To do this, n t trajectories, corresponding to the paths of a single agent, are generated and used as inputs to RNN models.…”

Section: Single Unit Representation Detailsmentioning

confidence: 99%

“…These RNN models [27,28], when trained on single agent path integration, have been shown to develop properties analogous to MEC. In particular, they exhibit units with grid, border, and spatially periodic "band" responses [27][28][29][30] and population dynamics consistent with continuous toroidal attractors [31,32]. While their biological plausibility remains unclear [33,34], we choose to analyze these RNN models for three reasons.…”

mentioning

confidence: 99%

See 3 more Smart Citations

Not so griddy: Internal representations of RNNs path integrating more than one agent

Redman,

Acosta,

Acosta–Mendoza

et al. 2024

Preprint

View full text Add to dashboard Cite

Success in collaborative and competitive environments, where agents must work with or against each other, requires individuals to encode the position and trajectory of themselves and others. Decades of neurophysiological experiments have shed light on how brain regions [e.g., medial entorhinal cortex (MEC), hippocampus] encode the self’s position and trajectory. However, it has only recently been discovered that MEC and hippocampus are modulated by the positions and trajectories of others. To understand how encoding spatial information of multiple agents shapes neural representations, we train a recurrent neural network (RNN) model that captures properties of MEC to path integrate trajectories of two agents simultaneously navigating the same environment. We find significant differences between these RNNs and those trained to path integrate only a single agent. At the individual unit level, RNNs trained to path integrate more than one agent develop weaker grid responses, stronger border responses, and tuning for therelativeposition of the two agents. At the population level, they develop more distributed and robust representations, with changes in network dynamics and manifold topology. Our results provide testable predictions and open new directions with which to study the neural computations supporting spatial navigation.

show abstract

Section: Resultssupporting

confidence: 92%

Section: Resultsmentioning

confidence: 99%

Section: Single Unit Representation Detailsmentioning

confidence: 99%

mentioning

confidence: 99%

See 2 more Smart Citations

Not so griddy: Internal representations of RNNs path integrating more than one agent

Redman,

Acosta,

Acosta–Mendoza

et al. 2024

Preprint

View full text Add to dashboard Cite

show abstract

“…In the context of grid cells, previous work has made use of clustering methods to identify cells whose activity plays a role in generating a torus [25] (the neural manifold regularly observed in grid cell recordings). Such cells have even been found to play a more fundamental role in encoding position in artificial recurrent neural network models performing a path integration task, than cells with a high grid-score [67].…”

Section: Discussionmentioning

confidence: 99%

Topological Structure of Population Activity in Mouse Visual Cortex Encodes Visual Scene Rotations

Beshkov

Einevoll

2023

Preprint

View full text Add to dashboard Cite

The primary visual cortex is one of the most well understood regions supporting the processing involved in sensory computation. Historically, our understanding of this part of the brain has been driven by describing the features to which individual neurons respond. An alternative approach, which is rapidly becoming a staple in neuroscience, is to study and analyze the geometry and topology of the manifold generated by the neural activity of large populations of neurons. In this work, we introduce a rigorous quantification of the structure of such neural manifolds and address some of the problems the community has to face when conducting topological data analysis on neural data. We do this by analyzing publicly available two-photon optical recordings of primary mouse visual cortex in response to visual stimuli with a densely sampled rotation angle. Since the set of two- dimensional rotations lives on a circle, one would hypothesize that they induce a circle-like manifold in neural activity. We confirm this hypothesis by discovering a circle-like neural manifold in the population activity of primary visual cortex. To achieve this, we applied a shortest-path (geodesic) approximation algorithm for computing the persistent homology groups of neural activity in response to visual stimuli. It is important to note that the manifold is highly curved and standard Euclidean approaches failed to recover the correct topology. Furthermore, we identify subpopulations of neurons which generate both circular and non-circular representations of the rotated stimuli, with the circular representations being better for angle decoding. We found that some of these subpopulations, made up of orientationally selective neurons, wrap the original set of rotations on itself which implies that the visual cortex also represents rotations up to 180 degrees. Given these results we propose that population activity can represent the angle of rotation of a visual scene, in analogy with how individual direction-selective neurons represent the angle of direction in local patches of the visual field. Finally, we discuss some of the obstacles to reliably retrieving the truthful topology generated by a neural population.

show abstract

Coherently Remapping Toroidal Cells but Not Grid Cells Are Responsible for Path Integration in Virtual Agents

Cited by 2 publications

References 40 publications

Not so griddy: Internal representations of RNNs path integrating more than one agent

Not so griddy: Internal representations of RNNs path integrating more than one agent

Topological Structure of Population Activity in Mouse Visual Cortex Encodes Visual Scene Rotations

Contact Info

Product

Resources

About