Remapping in a recurrent neural network model of navigation and context inference

Low, Isabel I. C.; Giocomo, Lisa M.; Williams, Alex H.

doi:10.7554/elife.86943

“…7 there) – that discrepancies between remapping probabilities can be explained as testing at different points of training. Beyond such abstract models, for future work, it is also possible to mechanistically model multiple environments and study remapping probabilities through them ( Low et al, 2023 ).…”

Section: Discussionmentioning

confidence: 99%

Representational drift as a result of implicit regularization

Ratzon,

Derdikman,

Barak

2024

eLife

1

0

View full text Add to dashboard Cite

Recent studies show that, even in constant environments, the tuning of single neurons changes over time in a variety of brain regions. This representational drift has been suggested to be a consequence of continuous learning under noise, but its properties are still not fully understood. To uncover the underlying mechanism, we trained an artificial network on a simplified navigational task, inspired by the predictive coding literature. The network quickly reached a state of high performance, and many neurons exhibited spatial tuning. We then continued training the network and noticed that the activity became sparser with time. We observed vastly different time scales between the initial learning and the ensuing sparsification. We verified the generality of this phenomenon across tasks, learning algorithms, and parameters. This sparseness is a manifestation of the movement within the solution space - the networks drift until they reach a flat loss landscape. This is consistent with recent experimental results demonstrating that CA1 neurons increase sparseness with exposure to the same environment and become more spatially informative. We conclude that learning is divided into three overlapping phases: Fast familiarity with the environment, slow implicit regularization, and a steady state of null drift. The variability in drift dynamics opens the possibility of inferring learning algorithms from observations of drift statistics.

show abstract

“…7 there) – that discrepancies between remapping probabilities can by explained as testing at different points of training. Beyond such abstract models, for future work, it is also possible to mechanistically model multiple environments and study remapping probabilities through them [44].…”

Section: Discussionmentioning

confidence: 99%

Representational drift as a result of implicit regularization

Ratzon

¹

,

Derdikman

²

,

Barak

³

2023

Preprint

0

View full text Add to dashboard Cite

Recent studies show that, even in constant environments, the tuning of single neurons changes over time in a variety of brain regions. This representational drift has been suggested to be a consequence of continuous learning under noise, but its properties are still not fully understood. To uncover the underlying mechanism, we trained an artificial network to perform a predictive coding task. After the loss converged, the activity slowly became sparser. We verified the generality of this phenomenon across modeling choices. This sparseness is a manifestation of drift in the solution space to a flatter area. It is consistent with recent experimental results demonstrating that CA1 spatial code becomes sparser after familiarity. We conclude that learning is divided into three overlapping phases: Fast familiarity with the environment, slow implicit regularization, and a steady state of null drift. These findings open the possibility of inferring learning algorithms from observations of drift statistics.

show abstract

“…Here, we test these two coding schemes by first using computational modeling on HD systems to explore algorithms that simultaneously encode the interdependent variables of HD and AHV, and then analyzing empirical data from mice's HD systems to verify the biological plausibility of these algorithms. Specifically, we first explored self-emergent algorithms through an encapsulated computational model specifically designed to achieve the computational goal of simultaneously encoding HD and AHV, where similar models in previous studies have revealed the conjunctive coding of task-related variables (Cueva et al, 2019;Low et al, 2023). This approach mitigates potential confounding effects from interactive influences of upstream and downstream cortical regions or from intrinsic functionalities not directly related to HD and AHV within biological HD systems.…”

Section: Introductionmentioning

confidence: 99%

“…Specifically, we first explored self-emergent algorithms through an encapsulated computational model specifically designed to achieve the computational goal of simultaneously encoding HD and AHV, where similar models in previous studies have revealed the conjunctive coding of task-related variables (Cueva et al, 2019; Low et al, 2023). This approach mitigates potential confounding effects from interactive influences of upstream and downstream cortical regions or from intrinsic functionalities not directly related to HD and AHV within biological HD systems.…”

Section: Introductionmentioning

confidence: 99%

Encoding of interdependent features of head direction and angular head velocity in navigation

Cai,

Liu,

Liu

2024

Preprint

2

0

View full text Add to dashboard Cite

To comprehend the complex world around us, our brains are tasked with the remarkable job of integrating multiple features into a cohesive whole. While previous studies have primarily focused on the processing and integration of independent features, here we investigated the simultaneous encoding of the interdependent features, specifically head direction (HD) and its temporal derivative, angular head velocity (AHV), by first employing computational modeling on HD systems to explore emergent algorithms and then validated its biological plausibility with empirical data from mice's HD systems. Our analysis revealed two distinct neuron populations: those with multiphasic tuning curves for HD compromised their HD encoding capacity to better capture AHV dynamics, while those with monophasic tuning curves primarily encoded HD. This pattern of functional dissociation was observed in both artificial HD systems and the cortical and subcortical regions upstream of biological HD systems, suggesting a general principle for encoding interdependent features. Further, exploration of the underlying mechanisms involved examining neural manifolds embedded within the representational space constructed by these neurons. We found that the manifold by neurons with multiphasic tuning curves was locally jagged and complex, which effectively expanded the dimensionality of the neural representation space and in turn facilitated a high-precision representation of AHV. Therefore, the encoding strategy for HD and AHV likely integrates characteristics of both dense and sparse coding schemes to achieve a balance between preserving specificity for individual features and maintaining their interdependency nature, marking a significant departure from the encoding of independent features and thus advocating future research delving into the encoding strategies of interdependent features.

show abstract

Remapping in a recurrent neural network model of navigation and context inference

Cited by 7 publications

References 57 publications

Representational drift as a result of implicit regularization

Representational drift as a result of implicit regularization

Representational drift as a result of implicit regularization

Encoding of interdependent features of head direction and angular head velocity in navigation

Contact Info

Product

Resources

About