Rationalizing constraints on the capacity for cognitive control

Trends in Cognitive Sciences, Год журнала: 2021, Номер 25(9), С. 757 - 775

Опубликована: Июль 28, 2021

Humans are remarkably limited in: (i) how many control-dependent tasks they can execute simultaneously, and (ii) intensely focus on a single task. These limitations universal assumptions of most theories cognition. Yet, rationale for why humans subject to these constraints remains elusive. This feature review draws recent insights from psychology, neuroscience, machine learning, suggest that cognitive control may result rational adaptation fundamental, computational dilemmas in neural architectures. The reviewed literature implies multitasking trade-off between learning efficacy processing efficiency the intensity commitment task reflect stability flexibility.

Язык: Английский

---

Cognitive Control as a Multivariate Optimization Problem

Journal of Cognitive Neuroscience, Год журнала: 2022, Номер 34(4), С. 569 - 591

Опубликована: Янв. 21, 2022

A hallmark of adaptation in humans and other animals is our ability to control how we think behave across different settings. Research has characterized the various forms cognitive can take-including enhancement goal-relevant information, suppression goal-irrelevant overall inhibition potential responses-and identified computations neural circuits that underpin this multitude types. Studies have also a wide range situations elicit adjustments allocation (e.g., those eliciting signals indicating an error or increased processing conflict), but rules governing when given situation will give rise adjustment remain poorly understood. Significant progress recently been made on front by casting as decision-making problem. This approach developed unifying normative models prescribe change incentives task demands result changes form control. Despite their successes, these models, experiments test them, yet face greatest challenge: deciding select among multiplicity configurations take at any time. Here, lay out complexities inverse problem inherent allocation, close parallels problems within motor choosing between redundant limb movements). We discuss existing solutions control's drawn from optimal theory, which proposed effort costs act regularize actions transform planning into well-posed These same principles may help shed light brains optimize over complex configuration, while providing new perspective origins mental effort.

Язык: Английский

---

Modelling continual learning in humans with Hebbian context gating and exponentially decaying task signals

PLoS Computational Biology, Год журнала: 2023, Номер 19(1), С. e1010808 - e1010808

Опубликована: Янв. 19, 2023

Humans can learn several tasks in succession with minimal mutual interference but perform more poorly when trained on multiple at once. The opposite is true for standard deep neural networks. Here, we propose novel computational constraints artificial networks, inspired by earlier work gating the primate prefrontal cortex, that capture cost of interleaved training and allow network to two sequence without forgetting. We augment stochastic gradient descent algorithmic motifs, so-called "sluggish" task units a Hebbian step strengthens connections between hidden encode task-relevant information. found introduce switch-cost during training, which biases representations under towards joint representation ignores contextual cue, while promotes formation scheme from layer produces orthogonal are perfectly guarded against interference. Validating model previously published human behavioural data revealed it matches performance participants who had been blocked or curricula, these differences were driven misestimation category boundary.

Язык: Английский

---

Continual task learning in natural and artificial agents

Trends in Neurosciences, Год журнала: 2023, Номер 46(3), С. 199 - 210

Опубликована: Янв. 20, 2023

How do humans and other animals learn new tasks? A wave of brain recording studies has investigated how neural representations change during task learning, with a focus on tasks can be acquired coded in ways that minimise mutual interference. We review recent work explored the geometry dimensionality neocortex, computational models have exploited these findings to understand may partition knowledge between tasks. discuss ideas from machine including those combine supervised unsupervised are helping neuroscientists natural learned biological brains.

Язык: Английский

---

Policy compression: An information bottleneck in action selection

˜The œPsychology of learning and motivation/˜The œpsychology of learning and motivation, Год журнала: 2021, Номер unknown, С. 195 - 232

Опубликована: Янв. 1, 2021

Язык: Английский

---

The role of conjunctive representations in prioritizing and selecting planned actions

eLife, Год журнала: 2022, Номер 11

Опубликована: Окт. 31, 2022

For flexible goal-directed behavior, prioritizing and selecting a specific action among multiple candidates are often important. Working memory has long been assumed to play role in prioritization planning, while bridging cross-temporal contingencies during selection. However, studies of working have mostly focused on for single components an plan, such as rule or stimulus, rather than management all these elements planning. Therefore, it is not known how post-encoding selection operate the entire profile representations prospective actions. Here, we assessed control processes unfold over representations, highlighting conjunctive that nonlinearly integrate task-relevant features maintenance plans. each trial, participants prepared two independent rule-based actions simultaneously, then they were retro-cued select one their response. Prior start was randomly assigned be high priority by cueing more likely tested. We found both full plans maintained preparation, regardless priority. output selection, representation high-priority plan enhanced readily selected output. Furthermore, strength associated with behavioral interference when low-priority Thus, alternate upcoming integrated served target attentional mechanisms prioritize from within memory.

Язык: Английский

---

Rationalizing Constraints on the Capacity for Cognitive Control

Опубликована: Ноя. 17, 2020

Humans are remarkably limited in (a) how many control-dependent tasks they can execute simultaneously, and (b) intensely focus on a single task. These limitations universal assumptions of most theories cognition. Yet, rationale for why humans subject to these constraints remains elusive. This review draws recent insights from psychology, neuroscience machine learning, suggest that cognitive control may result rational adaptation fundamental computational dilemmas neural architectures. The reviewed literature implies multitasking tradeoff between learning efficacy processing efficiency, the intensity commitment task reflect stability flexibility.

Язык: Английский

---

An Integrated Model of Semantics and Control

Опубликована: Июнь 26, 2023

Understanding the mechanisms enabling learning and flexible use of knowledge in context-appropriate ways has been a major focus research study both semantic cognition cognitive control. We present unified model semantics control that addresses these questions from perspectives. The provides coherent view how knowledge, ability to flexibly access deploy meet current task demands, arises end-to-end statistics environment. show unresolved issues literatures, including operates over features covary with one another representations themselves are structured emerge through learning, series behavioral experiments simulations. conclude by discussing implications our approach other fundamental science, machine artificial intelligence.

Язык: Английский

---

Reconciling shared versus context-specific information in a neural network model of latent causes

Scientific Reports, Год журнала: 2024, Номер 14(1)

Опубликована: Июль 22, 2024

Abstract It has been proposed that, when processing a stream of events, humans divide their experiences in terms inferred latent causes (LCs) to support context-dependent learning. However, shared structure is present across contexts, it still unclear how the “splitting” LCs and learning can be simultaneously achieved. Here, we Latent Cause Network (LCNet), neural network model LC inference. Through learning, naturally stores that tasks weights. Additionally, represents context-specific using context module, controlled by Bayesian nonparametric inference algorithm, which assigns unique vector for each LC. Across three simulations, found LCNet could (1) extract function task while avoiding catastrophic interference, (2) capture human data on curriculum effects schema (3) infer underlying event naturalistic videos daily events. Overall, these results demonstrate computationally feasible approach reconciling scalable from laboratory experiment settings settings.

Язык: Английский

---

A Transient High-dimensional Geometry Affords Stable Conjunctive Subspaces for Efficient Action Selection

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2023, Номер unknown

Опубликована: Июнь 11, 2023

Abstract Flexible action selection requires cognitive control mechanisms capable of mapping the same inputs to different output actions depending on context. From a neural state-space perspective, this representation that separates similar input states by Additionally, for be robust and time-invariant, information must stable in time, enabling efficient readout. Here, using EEG decoding methods, we investigate how geometry dynamics representations constrain flexible human brain. Participants performed context-dependent task. A forced response procedure probed trajectories. The result shows before successful responses, there is transient expansion representational dimensionality separated conjunctive subspaces. Further, stabilizes time window, with entry into stable, high-dimensional state predictive individual trial performance. These results establish brain needs over behavior.

Язык: Английский

---