nucleation

LIT-1195: Attention or salience?

2 min read

Properties5
StatusActive — read 2026-10-16 (NOTE-956) in full, from the authors' accepted manuscript on UCL Discovery; worth reading as the short statement, within active inference, that attentional gain (the estimated precision of currently available data) and salience (the expected information gain of an action that would acquire new data, usually a saccade) are distinct computations with distinct disorders, hallucination and neglect, and that they are coupled because a saccade is a prediction of high precision at the location foveated. Read it knowing that it is a five-page review with no new data or simulations; that it assigns covert spatial attention to gain by stipulation; that it never treats gain as an action; and that it ends by endorsing the premotor theory, so covert attention changes what is observed only by biasing overt saccades.
Version2
Tagsneuroscience, probabilistic-modeling, cognition
SupportsCLAIM-337 — Active inference already runs an internal loop: mental action selects among higher-level cognitive states and sets the precision of lower-level observations, structure learning reshapes the generative model without action, and a judgement is behaviour, so a judgement–observation loop may be action and perception one Markov blanket inward
Cited as “paper” byNOTE-956 — Parr & Friston — attention or salience?

Graph View

Parr, T. & Friston, K. J. (2019). Current Opinion in Psychology 29:1–5. doi:10.1016/j.copsyc.2018.10.006. Accepted manuscript: UCL Discovery, eprint 10058428.

Key takeaways

From the reading, NOTE-956. Page numbers are the accepted manuscript’s.

  • Two selections. Attention “depends upon estimating the precision of sensory data”, salience “is a consequence of the need to actively engage with the sensorium” (abstract, p. 1). One weights current data, the other acquires new data (p. 5).
  • Salience is information gain. “Salient actions are those that solicit unambiguous data that could not have been predicted with a high degree of certainty prior to that action” (p. 4).
  • The coupling. A saccade command is a prediction of high precision at the foveated location, so “deployment of covert attention (precision) to a given location biases planning in favour of an overt saccade to that (salient) location” (p. 5).
  • Disorders. Gain failures give “false percepts, including hallucinations”; neglect is “a disorder of salience attribution” (pp. 4–5).
  • Scope. A review with no new data. Covert spatial attention is grouped with gain by stipulation. Gain is never treated as an action.

Why this is here

Registered for CLAIM-337, on the standard reading that attention weights data and salience belongs to action. The reading confirms the split but not the use. Attentional gain here is inference about precision, not a mental action. It changes how much a datum counts, not what the datum is, and the content is fixed by the world and the sensors, not, as the registration said, by the generative model. Its one route to changing what is observed is through overt saccades, and the paper ends by endorsing the premotor theory. So it gives CLAIM-341’s line between weighting and producing an observation a clean statement, without supplying an internal loop on either side.

Salience bears on CLAIM-339 and CLAIM-342. It makes action a seeker of data the agent cannot predict, which the counter needs. But it scores candidate actions by the agent’s own precision beliefs, and the streetlight effect follows, which is the anchor claim’s warning.

Backlinks