Geometry of neural computation unifies working memory and planning

Abstract

Real-world tasks require coordination of working memory, decision making, and planning, yet these cognitive functions have disproportionately been studied as independent modular processes in the brain. Here we propose that contingency representations, defined as mappings for how future behaviors depend on upcoming events, can unify working memory and planning computations. We designed a task capable of disambiguating distinct types of representations. Our experiments revealed that human behavior is consistent with contingency representations, and not with traditional sensory models of working memory. In task-optimized recurrent neural networks we investigated possible circuit mechanisms for contingency representations and found that these representations can explain neurophysiological observations from prefrontal cortex during working memory tasks. Finally, we generated falsifiable predictions for neural data to identify contingency representations in neural data and to dissociate different models of working memory. Our findings characterize a neural representational strategy that can unify working memory, planning, and context-dependent decision making.

Prioritization of semantic over visuo-perceptual aspects in multi-item working memory

“…In summary, our results add to ongoing discussions about the putative nature of “working” memories, including the extent to which they are retrospective (in terms of a sustained ‘copy’ of past sensory inputs) or prospective (in terms of an emerging action plan or response contingency 43,61 ) or both, and how they are distinguished—or not— from long-term memories. Our study provides new evidence for LTM-like dynamics in a typical WM task context (sub-span multi-item maintenance).…”

Constructing neural network models from brain data reveals representational transformations underlying adaptive behavior

“…Instead, our conjunctive representations are consistent with the biased competition theory of attention, where additive computations were passed through a nonlinearity 32 . It will be important for future work to distinguish the content of these two types of stimulus and rule conjunctions, and whether the representational content of conjunctions simultaneously contain both stimulus and rule content, or if instead rule and stimulus conjunctions collapse into contingency states for action selection as previously reported in computational models 38 . Nevertheless, these findings provide evidence to fill an important gap within the Flexible Hub framework, suggesting that the flexibility of rule updates are useful insofar as they can be integrated to form conjunctions with stimulus activity.…”

Ketamine induces multiple individually distinct whole-brain functional connectivity signatures

“…For the LSD and psilocybin analyses we used the Δ GBC parcellated neural maps from two independent pharmacological neuroimaging datasets (51, 53). We calculated effective dimensionality for each drug, using a re-sampling method to ensure that sample size (which differed between the three datasets) was kept constant and did not bias the results (see Methods) (54–56). This revealed that ketamine was significantly higher in dimensionality than LSD and psilocybin ( Fig.…”

“…Effective dimensionality was calculated on the ketamine, psilocybin, and LSD datasets to compare the dimensionality of the neural effects of different pharmacological substances. We used the participation ratio (PR), calculated as: where {λ i } is the ith eigenvalue of the covariance matrix, and (54, 56). Larger values indicate a more complex higher dimensional dataset, while smaller values indicate a less complex lower dimensional dataset.…”