Contribution
A conceptual review, with no new data or simulations. It argues that the two senses of “attention” in the literature are different computations with different anatomy and different disorders. One is covert gain control over sensory channels. The other is the selection of overt sampling actions by their epistemic value. It then says how the two are coupled. Its formal backing is cited, chiefly to Parr and Friston (2017, Scientific Reports), which “sets out the distinction between attention and salience formally (mathematically) and illustrates this through simulation” (reference 5 annotation, p. 6).
Key insight
“Attentional processes select those sensations that are informative about their causes. Broadly, these processes can be separated into two categories – those that optimise the weighting of current data, and those that mediate the acquisition of new sensations” (p. 5). Attention works on data already in hand; salience works on data not yet acquired. “The former relies upon currently available data, while the latter depends on data that has yet to be acquired” (p. 4).
Assumptions
- A generative model. Perception is “a process of optimising beliefs about the causes of sensations, while actions select new sensory data to test hypotheses about these causes” (pp. 1–2).
- Beliefs are subpersonal. “These are not propositional or subjective beliefs and are generally considered subpersonal” (footnote 1, p. 2).
- Precision belongs to the likelihood. “the degree to which we should update our beliefs about the causes of a sensation depends upon the precision (inverse covariance) of the probabilistic mapping between the two … precision corresponds to the confidence afforded sensory evidence, given our knowledge about how it was generated” (p. 3).
- Covert spatial attention is gain. “For the purposes of this paper, we group visuospatial covert attention (in which no eye-movement takes place) with the covert attentional processes considered in the previous section” (p. 4). That is a stipulation, made against the premotor theory’s reading of covert attention as a planned saccade.
Key results
The paper’s results are conceptual and rest on cited work.
- Attention as gain control. Implemented as “a multiplicative gain on ascending (‘bottom-up’) connections” (p. 3). It reproduces Posner cueing and figure-ground segregation by second-order statistics. Acetylcholine is the leading candidate for the gain, with Lewy body hallucinations as the pathology (p. 3).
- Salience. “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). Salience is equivalently intrinsic motivation, Bayesian surprise, relative entropy, or the epistemic part of expected free energy, and has the form of an optimal experimental design criterion (p. 4). It explains inhibition of return. Visual neglect is read as “a disorder of salience attribution” (p. 4), with the basal ganglia evaluating candidate saccades.
- The coupling. The streetlight effect is optimal in a sense, since “the unambiguous (precise) visual data obtained under the light supports stronger inferences than the darkness around it” (pp. 4–5). And because motor commands are predictions, “a prediction (i.e., motor command or setpoint for oculomotor reflexes) that a location in visual space will be foveated entails a belief that the precision associated with this location is high. The resulting saccade fulfils this prediction” (p. 5). Subthreshold frontal-eye-field stimulation raises gain at the location that suprathreshold stimulation would saccade to, “exactly as if the precision had increased” (p. 5). So “deployment of covert attention (precision) to a given location biases planning in favour of an overt saccade to that (salient) location” (p. 5).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Gain control and salience are distinct computations, with distinct anatomy and disorders (hallucination versus neglect) | moderate | review of cited theory and lesion evidence |
| C2 | Attentional gain is the estimated precision of the likelihood of current data | moderate | cited models (Feldman and Friston 2010 and others) |
| C3 | Salience is the expected information gain of a sampling action | moderate | cited formal equivalences (Itti and Baldi, Lindley, expected free energy) |
| C4 | Covert attention to a location biases an overt saccade there, because a saccade is a prediction of high precision at that location | weak | argument plus one stimulation study (Moore and Fallah 2001) |
Concepts
- Attention (gain control). Covert weighting of current data by their estimated precision. It changes how much data count, not which data arrive.
- Salience. A property of an action: how much the data it would bring would change beliefs. High when the outcome is uncertain but the mapping from causes to outcome is precise (p. 4).
- Precision. Inverse covariance of the likelihood mapping from causes to sensations.
Connections
It summarizes the authors’ earlier work (Sci Rep 2017; J R Soc Interface 2017; Cereb Cortex 2017 on neglect), Feldman and Friston (2010) on attention as precision, and the epistemic-value formulation of Friston et al. (2015) and Mirza et al. (2016). Its foil is the premotor theory of attention (Rizzolatti et al. 1987), which it ends by endorsing in modified form. It is the source Sandved-Smith et al. (LIT-1191, read in NOTE-958) build their attentional level on. It does not use the term “Markov blanket” and does not speak of mental action.
The LIT’s summary and “Why this is here”. The summary is right in substance: attention “depends upon estimating the precision of sensory data”, while salience “is a consequence of the need to actively engage with the sensorium” (abstract, p. 1). Three corrections.
- The split is not the paper’s last word. Its last section reconciles them: both “represent beliefs about the capacity of sensory data to adjudicate among competing hypotheses” (p. 4). And in the spatial domain a precision belief and a saccade command are the same prediction (p. 5).
- Attention here is not an internal action. The paper never calls gain control an action or a choice. It is inference about precision. What it does to perception is weight data whose content attention does not set. So the registration’s “attention weights data whose content the generative model already fixes” is half right. The content is fixed by the world and the sensors, not by the model. What the model fixes is the mapping whose precision is estimated.
- Attention can change what is observed, through action. Covert precision on a location “biases planning in favour of an overt saccade” (p. 5), and the saccade then changes the data. So on this paper’s account attention changes what is observed only by recruiting overt action.
CLAIM-337 (the counter). The claim cites this paper, “on the standard reading”, for the split between attention and salience. The reading confirms the split, but it does not help the claim’s argument for an internal loop. Gain control here is not selected as a policy and is not called mental action. Its one route to changing what is observed runs through the eyes. If anything the paper pulls covert attention back toward overt action: its conclusion endorses the premotor theory. The claim’s first strand rests on Sandved-Smith et al., and should cite this paper only for the identification of attention with likelihood precision.
CLAIM-341 (the thesis). Where the thesis locates its disagreement, the paper sits on the thesis’s side of the line for attention. Attentional gain changes how much a fixed datum counts, not what the datum is. But that is not the thesis’s loop either. The paper’s gain is a rate on updating, and nothing in it reshapes the observer over time. On salience the paper is the clearest statement of the view the thesis rejects: the agent’s influence on what it observes is exercised by choosing where to look.
CLAIM-339, CLAIM-342 and QUESTION-121. Salience bears on both sides of this exchange. Against Friston et al. (2010, read in NOTE-960), where action samples what the agent expects, salient action seeks data “that could not have been predicted with a high degree of certainty” (p. 4): “the best (most salient) experiments are those that bring about the greatest change in beliefs” (p. 4). That is action as a route to evidence the current interpretation cannot foresee, which is what CLAIM-339 says action supplies. But the agent still scores salience with its own current beliefs about precision, and the streetlight effect is the result: it looks where its model says data will be precise (pp. 4–5). Evidence chosen by epistemic value is thus routed through the interpretation it tests, which is CLAIM-342’s warning. Gain control also has its own anchor problem. Its failure, weighting priors over evidence, gives “false percepts, including hallucinations” (p. 5). That failure involves no action at all.
Bearing on the record
- CLAIM-337: the paper confirms that attention is likelihood precision and salience is the epistemic value of action. It does not supply an internal loop, and it ties covert attention to overt sampling.
- CLAIM-341: attention weights a fixed observation, as the claim’s line requires, but does not reshape the observer.
- CLAIM-339 and CLAIM-342: salience makes action a seeker of unpredicted data, which the counter needs, while scoring it by the agent’s own precision beliefs, as the anchor claim warns.
- No instruction for machine-learning practice.
Limitations
It is a five-page opinion review. Its claims are argued from cited models and from lesion, pharmacological and stimulation studies, not shown here. Covert spatial attention is assigned to gain control by stipulation. The coupling argument (p. 5) rests on one stimulation result and on the premise that motor commands are predictions.
Open questions
- Is there a covert analogue of salience? That would be choosing which internal channel to weight by the information it would yield, without moving. The paper groups covert spatial attention with gain and so does not ask.
- Does the streetlight effect, where salience is scored by current precision beliefs, lock in under a wrong precision model? NOTE-766’s sensory-attenuation lock-in suggests it can.