The Geometry of Abstraction in the Hippocampus and Prefrontal Cortex

Bernardi, Silvia; Benna, Marcus K.; Rigotti, Mattia; Munuera, Jérôme; Fusi, Stefano; Salzman, C. Daniel

doi:10.1016/j.cell.2020.09.031

Cited by 281 publications

(524 citation statements)

References 52 publications

Supporting

Mentioning

417

Contrasting

Order By: Relevance

“…Consistent with the morphology of the SourceAll response decoding timecourse (Figure 4), we again observed two decoding peaks across networks, which were linked via a supplementary temporal generalization analysis to distinct multivariate codes, possibly reflecting motor preparation and motor execution/feedback processes (see SI, Figure S2). Response information was broadly decodable across networks, consistent with prior findings from primate electrophysiology (Bernardi et al, 2020; Siegel et al, 2015) and rodent optical imaging (Kauvar et al, 2020).…”

Section: Resultssupporting

confidence: 82%

Network modeling of dynamic brain interactions predicts emergence of neural information that supports human cognitive behavior

Mill

Hamilton

Winfield

et al. 2021

Preprint

View full text Add to dashboard Cite

SummaryHow cognitive task information emerges from brain activity is a central question in neuroscience. We identified the spatiotemporal emergence of task information in the human brain using individualized source electroencephalography and dynamic multivariate pattern analysis. We then substantially extended recently developed brain activity flow models to predict the future emergence of task information dynamics. The model simulated the flow of task-evoked activity over causally interpretable resting-state functional connections (dynamic, lagged, direct and directional) to accurately predict response information dynamics underlying cognitive task behavior. Predicting event-related spatiotemporal activity patterns and fine-grained representational geometry confirmed the model’s faithfulness to how the brain veridically represents response information. Simulated network “lesioning” revealed cognitive control networks (CCNs) as the dominant causal drivers of response information flow. These results demonstrate the efficacy of dynamic activity flow models in predicting the emergence of task information, thereby revealing a mechanistic role for CCNs in producing behavior.HighlightsNetwork models fit with EEG connectivity simulated the emergence of task informationThese activity flow models yield causal insights by predicting future neural activityThe models accurately predicted empirical response information underlying behaviorThe findings clarify fundamental principles of network computation in the human braineTOC blurb (In Brief)The authors comprehensively map the spatiotemporal signatures of task information in the brain, and then use a novel network modeling approach to predict future information dynamics underlying behavior. The findings link functional connectivity to a role in mediating information flow, highlighting cognitive control networks as causal drivers of this process.

show abstract

Section: Resultssupporting

confidence: 82%

Network modeling of dynamic brain interactions predicts emergence of neural information that supports human cognitive behavior

Mill

Hamilton

Winfield

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Similarly, when humans learned to rank images of animals according to their probability of paying out a reward, shared multivariate patterns in EEG signals come to code for the number and value, as if there were a corresponding neural signal for higher numbers and higher event probabilities (Luyckx et al, 2019). Together, these findings imply a general principle whereby neural state space alignment permits generalisation across contexts (Bernardi et al, 2020). However, generalisation of relational information between contexts can be hampered by what we call the "mapping problem": the need to represent stimulus geometry on a common scale.…”

Section: Discussionmentioning

confidence: 92%

“…Secondly, the fact that number lines are parallel in neural state space facilitates generalisation between physically similar stimuli that share a common magnitude. Irrespective of whether individual neurons exhibit specialised coding or exhibit mixed selectivity, this neural geometry ensures that a linear decoder trained to classify stimuli according to their magnitude in one context could be successfully applied to read out magnitudes in another, thus permitting crosscontext abstraction (Bernardi et al, 2020). Indeed, a long tradition emphasises that humans generalise naturally between space, time and number, by using a magnitude representation that is shared across different input modalities (Hubbard et al, 2005;Walsh, 2003).…”

Section: Discussionmentioning

confidence: 99%

“…One view is that conceptual knowledge relies on neural ensembles that code for relations among stimuli but are invariant to their physical properties (Baram et al, 2019;Behrens et al, 2018;Bellmund et al, 2018;Collins & Frank, 2013;Doumas et al, 2008;Brenden M. Lake et al, 2015;Summerfield et al, 2019;Tervo et al, 2016). Recent evidence hints that when sets of stimuli share relational structure across contexts, they are embedded on parallel low-dimensional neural manifolds, so that a linear decoder learned in one context can be readily repurposed for another (Bernardi et al, 2020;Fitzgerald et al, 2013;Ganguli et al, 2008;Luyckx et al, 2019;Remington et al, 2018). By aligning neural state spaces between contexts in this way, one can generalise relational knowledge, for example applying a criterion that distinguishes fast and slow animals to discriminate fast and slow vehicles, such as space rockets and bicycles (Fig.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Neural state space alignment for magnitude generalisation in humans and recurrent networks

Sheahan

Luyckx

Nelli

et al. 2020

Preprint

View full text Add to dashboard Cite

A prerequisite for intelligent behaviour is to understand how stimuli are related and to generalise this knowledge across contexts. Generalisation can be challenging when relational patterns are shared across contexts but exist on different physical scales. Here, we studied neural representations in humans and recurrent neural networks performing a magnitude comparison task, for which it was advantageous to generalise concepts of “more” or “less” between contexts. Using multivariate analysis of human brain signals and of neural network hidden unit activity, we observed that both systems developed parallel neural “number lines” for each context. In both model systems, these number state spaces were aligned in a way that explicitly facilitated generalisation of relational concepts (more and less). These findings suggest a previously overlooked role for neural normalisation in supporting transfer of a simple form of abstract relational knowledge (magnitude) in humans and machine learning systems.

show abstract

“…A possible explanation for this diversity is that different subregions of the vmPFC serve different functions. Hence, a strong possibility is that the vmPFC represents the relevant cognitive map for the current task (Bernardi et al, 2018; Schuck, Cai, Wilson, & Niv, 2016; Wilson, Takahashi, Schoenbaum, & Niv, 2014), and therefore the nature of the coding scheme in vmPFC may depend on the demands of the task at hand. A decision-maker does not necessarily need to know the spatial relationships between different choice options in attribute space but does need to know the relevant ordering of different options in terms of her preferences.…”

Section: Discussionmentioning

confidence: 99%

Subjective value, not a gridlike code, describes neural activity in ventromedial prefrontal cortex during value-based decision-making

Lerman

et al. 2019

Preprint

View full text Add to dashboard Cite

Across numerous studies, ventromedial prefrontal cortex (vmPFC) activity has been found to correlate with subjective value during value-based decision-making. Recently, however, vmPFC has also been shown to reflect a grid-like code used for navigation through physical and conceptual space. This raises the possibility that such a grid-like code could broadly underlie vmPFC's role in complex cognition and, specifically, may be used for comparing options in decision-making. Here, we first show that, in theory, a grid-like code could account for vmPFC activity previously attributed to subjective value. We then test for grid-like modulation in a large fMRI dataset of intertemporal choice. Here, however, we find that the grid-like model fails confirmatory tests and that instead subjective value provides the better description for vmPFC responses. Our results constrain the tasks for which grid-like modulation is observed in vmPFC and suggest that vmPFC might implement the most useful code for the current task.

show abstract

The Geometry of Abstraction in the Hippocampus and Prefrontal Cortex

Cited by 281 publications

References 52 publications

Network modeling of dynamic brain interactions predicts emergence of neural information that supports human cognitive behavior

Network modeling of dynamic brain interactions predicts emergence of neural information that supports human cognitive behavior

Neural state space alignment for magnitude generalisation in humans and recurrent networks

Subjective value, not a gridlike code, describes neural activity in ventromedial prefrontal cortex during value-based decision-making

Contact Info

Product

Resources

About