Task representations in neural networks trained to perform many cognitive tasks

Yang, Guangyu Robert; Joglekar, Madhura R.; Song, Hui; Newsome, William T.; Wang, Xiao Jing

doi:10.1038/s41593-018-0310-2

Cited by 416 publications

(485 citation statements)

References 38 publications

Supporting

Mentioning

429

Contrasting

Order By: Relevance

“…Artificial neural network models (ANNs) have been shown to replicate several properties observed in the LPFC during working memory tasks (Compte et al, 2000;Wimmer et al, 2014;Murray et al, 2017;Yang et al, 2019). However, the observation of the existence of a stable subspace in the presence of code-morphing imposes new constraints on these models.…”

Section: Fig 1 | Experimental Design and Code-morphing A Behavioramentioning

confidence: 99%

See 1 more Smart Citation

Time-Invariant Working Memory Representations in the Presence of Code-Morphing in the Lateral Prefrontal Cortex

Parthasarathy

Cheng

Herikstad

et al. 2019

Preprint

View full text Add to dashboard Cite

Endogenous processes allow the maintenance of working memories. These processes presumably involve prefrontal networks with strong recurrent connections. Distractors evoke a morphing of the population code, even when memories are stable. But it is unclear whether these dynamic population responses contain stable memory information. Here we show that dynamic prefrontal activity contains stable memory information, and the stability depends on parallel movement of trajectories associated with different memories in state space. We used an optimization algorithm to find a subspace with stable memory information. In correct trials the stability extended to periods that were not used to find the subspace, but in error trials the information and the stability were reduced. A bump attractor model was able to replicate these behaviors. The model provided predictions that could be confirmed with the neural data. We conclude that downstream regions could read memory information from a stable subspace.

show abstract

Section: Fig 1 | Experimental Design and Code-morphing A Behavioramentioning

confidence: 99%

“…Network models have been shown to replicate several properties of LPFC activity, including code stability (Compte et al, 2000;Wimmer et al, 2014;Murray et al, 2017;Yang et al, 2019). However, it is not known whether these models can exhibit code-morphing while retaining a stable subspace.…”

Section: Introductionmentioning

confidence: 99%

Time-Invariant Working Memory Representations in the Presence of Code-Morphing in the Lateral Prefrontal Cortex

Parthasarathy

Cheng

Herikstad

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

“…However, even if we happened upon a special, albeit common, scenario using these two tasks, it 373 is remarkable to observe a situation in which the large attention-related change in behavioral 374 performance can be accomplished without changing information coding or weights between 375 areas. In contrast, theoretical models and machine learning techniques accomplish flexibility in 376 computation almost solely by changing weights [44][45][46][47] . Our results constitute an existence proof: an 377 example of a situation in which flexibility can be mediated by simple changes within sensory 378 cortex.…”

mentioning

confidence: 99%

Simultaneous multi-area recordings suggest that attention improves performance by reshaping stimulus representations

Ruff

Cohen

2018

Preprint

View full text Add to dashboard Cite

6Abstract -Visual attention dramatically improves subjects' ability to see and also modulates the 7 responses of visual and oculomotor neurons. Despite hundreds of studies demonstrating the co-8 occurrence of behavioral and neuronal effects of attention 1 , the relationship between neuronal 9 modulations and improved performance remains unknown. There are three dominant hypotheses 10 was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.

show abstract

“…Although reinforcement learning has provided invaluable insight into the mechanism the PFC may use, it remains unclear how the PFC is able to encode multiple schemas, building on each other, without interference, and persisting so they may be accessed again in the future. The majority of models capable of solving multi-strategy problems require specially curated training regimens, most often by interleaving examples of different problem types [10]. Models learn successfully due to the balanced presentation of examples in training; if the training regimen is altered-for example, problem types appear in sequence rather than interleaved, as often happens in the world-the unbalanced models fail miserably [11].…”

Section: Introductionmentioning

confidence: 99%

A modeling framework for adaptive lifelong learning with transfer and savings through gating in the prefrontal cortex

Tsuda

Tye

Siegelmann

et al. 2020

Preprint

View full text Add to dashboard Cite

The prefrontal cortex encodes and stores numerous, often disparate, schemas and flexibly switches between them. Recent research on artificial neural networks trained by reinforcement learning has made it possible to model fundamental processes underlying schema encoding and storage. Yet how the brain is able to create new schemas while preserving and utilizing old schemas remains unclear. Here we propose a simple neural network framework based on a modification of the mixture of experts architecture to model the prefrontal cortex's ability to flexibly encode and use multiple disparate schemas. We show how incorporation of gating naturally leads to transfer learning and robust memory savings. We then show how phenotypic impairments observed in patients with prefrontal damage are mimicked by lesions of our network. Our architecture, which we call DynaMoE, provides a fundamental framework for how the prefrontal cortex may handle the abundance of schemas necessary to navigate the real world.

show abstract

Task representations in neural networks trained to perform many cognitive tasks

Cited by 416 publications

References 38 publications

Time-Invariant Working Memory Representations in the Presence of Code-Morphing in the Lateral Prefrontal Cortex

Time-Invariant Working Memory Representations in the Presence of Code-Morphing in the Lateral Prefrontal Cortex

Simultaneous multi-area recordings suggest that attention improves performance by reshaping stimulus representations

A modeling framework for adaptive lifelong learning with transfer and savings through gating in the prefrontal cortex

Contact Info

Product

Resources

About