Back

A plastic attractor model of flexible rule-based selective attention

Whyte, C. J.; Manohar, S. G.; Feredoes, E.; Woolgar, A.

2025-09-11 animal behavior and cognition
10.1101/2025.09.07.674747 bioRxiv
Show abstract

A defining feature of human cognition is the ability to select and respond to stimuli flexibly in different circumstances. Here we show that a recently proposed class of small associative neural network, plastic attractors, can perform such flexible cognitive functions through the rapid formation of task-based attractors. We simulated a rule-based selective attention paradigm, in which agents respond to one task-relevant feature of a visual stimulus, while ignoring another, irrelevant feature, and respond to the attended feature according to a predefined stimulus-response mapping rule. The model consists of a broadly tuned prefrontal population with rapidly changing recurrent connections to sensory neurons that compete via lateral inhibition. In this framework, the rules governing the focus of selective attention are not coded explicitly, but arise as an emergent property of temporary associations between stimulus features and motor responses. The model exhibited activation properties that embody cornerstone concepts in current attentional theory including mixed selectivity, adaptive coding and biased competition, and reproduced a number of classic behavioral and neural findings. A causal test of the model using non-invasive brain stimulation concurrent with functional magnetic resonance imaging (TMS-fMRI) in humans showed that network perturbation reproduced neural decoding and behavioural data. When features were task-relevant, they rapidly formed synaptic connections with frontal cortex binding them into an attracting state, which manifest as prioritized representation of attended information, but this state was readily corrupted by brain stimulation. The model shows mechanistically that rapid synaptic changes could explain flexible rule-based control of attention.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.