Salience network
The functional-network construct that recurs across the wiki’s predictive and anatomical accounts. Anchored on the anterior insula and ACC (with amygdala and hypothalamus), identified by Seeley et al. (2007) as an emotional “salience” network dissociable from an “executive-control” network (DLPFC + parietal). See craig-2009-anterior-insula, ad-craig.
Role in the two frameworks
- Craig (2009): cites the salience network as evidence for the AIC/ACC dyad and notes that the right AIC appears to switch between the central-executive and default-mode networks — a hub role consistent with the AIC as the apex of the global-emotional-moment.
- Seth (2013): locates interoceptive-inference within a salience network “anchored on the anterior insular and anterior cingulate cortices,” where interoceptive predictions are generated, compared, and updated (his Figure 2). So the same network is read as an ascending salience detector (Craig) and as a predictive comparator (Seth) — the two poles of feedforward-vs-predictive-interoception.
Overlap with core affect and the VMAs
The salience network’s core (AIC, ACC, amygdala) substantially overlaps both Lindquist et al.’s (2012) empirically derived core affect network and Seth & Friston’s (2016) anatomically defined visceromotor-areas (AIC, ACC, SGC, OFC) — three independent framings (functional-network, meta-analytic, cytoarchitectural) converging on nearly the same region set.
The insular anchor, isolated at rest (Cauda et al. 2011)
Seeley et al. (2007) defined the network from an anterior-insula/ACC co-activation, treating “the anterior insula” as its insular anchor. Cauda et al. (2011) show, from resting-state connectivity alone, that the anchor is more specific than that: only the ventral-anterior insula recovers the rACC-anchored limbic/salience pattern, while the dorsal-posterior insula belongs to a separate, anticorrelated sensorimotor network. So the salience network’s insular node is one pole of a within-insula dichotomy, not “the insula” — a distinction the network literature often glosses. Two consequences for this page. It supplies the network’s insular end with an intrinsic-connectivity grounding (the ventral-anterior seed, at rest, without a salient stimulus to detect — which sits awkwardly with the pure “detects” reading and comfortably with Barrett’s “assigns” and the switching accounts, since an intrinsic network needs no external salient event to be active). And its lateralization — the salience/anterior pattern more strongly coupled to rACC on the right, and right anterior ROIs preferentially wired to brainstem, pons and right thalamus — is early resting-state support for the right-insula-as-arousal/attention-hub reading (Craig 2009; Sridharan et al. 2008) that recurs across insular-cortex and Haruki. It also anticipates Chen et al.’s ventral/dorsal split of the network below: Cauda’s dichotomy is drawn along the insula, but it is the same warning that “the salience network” is internally divided.
The insular anchor splits again — and the dorsal node is specifically executive (Chang et al. 2013)
Cauda isolates the ventral-anterior insula as the salience/limbic pole. Chang et al. (2013) split the anterior insula once more and, via reverse-inference decoding, put a specific function on each half: the ventroanterior node is affective-chemosensory (amygdala/OFC-coupled), the dorsoanterior node executive (ACC/DLPFC-coupled). That is the within-insula version of the ventral/dorsal division Chen et al. flag across the whole network below (a homologous ventral/affective arm, a human-elaborated dorsal/attentional one) — and it lands on the same seam Cauda’s lateralization and Haruki’s cardiac/arousal cluster do.
It also supplies a caution the “switching/detects” readings need. By forward inference the dorsoanterior insula is the network’s most consistently activated node across nearly all tasks — which is exactly what makes it look like a general-purpose salience/control hub. Chang show that by reverse inference it is nonetheless specifically executive: its breadth is high consistency, not low specificity. So “the salience network activates for everything” is a forward-inference fact about its dorsoanterior anchor, not evidence that it does one domain-general thing — the same consistency-vs-specificity distinction the locationist-vs-constructionist-brain-emotion debate turns on.
A dissent about the amygdala node
LeDoux (2012) does not use the salience-network construct, but his argument bears directly on one of its nodes. Where Lindquist et al. reframe the amygdala as a domain-general salience detector — responsive to novel, uncertain, arousing, or salient stimuli of any valence — LeDoux holds that appetitive and aversive signals are processed by distinct neuronal populations in lateral/basal amygdala (Paton et al. 2006; Belova et al. 2007), and that “the fact that both appetitive and aversive stimuli activate the amygdala in fMRI studies… does not mean that these stimuli are processed the same by the amygdala.”
So “salience detector” may be an artifact of resolution: what looks domain-general at voxel scale could be several function-specific circuits sharing an address. This does not refute the network construct — which is defined by connectivity, not by claims about cellular homogeneity — but it does caution against reading a node’s breadth of activation as evidence that it performs one domain-general operation. See locationist-vs-constructionist-brain-emotion.
A third role: it doesn’t detect salience, it assigns it (Barrett 2017)
Barrett (2017) gives the network a job that is neither Craig’s nor Seth’s, and it is the most specific computational hypothesis on this page.
I further hypothesize that the salience network tunes the internal model by predicting which prediction errors to pay attention to — those errors that are likely to be allostatically relevant and therefore worth the cost of encoding and consolidation; called precision signals.
The mechanism is gain control, and it runs before the error arrives: precision signals “directly alter the gain on neurons that compute prediction error from incoming sensory input (i.e. they apply attention),” signalling confidence in the priors, confidence in the incoming signal’s reliability, and predicted relevance for allostasis. Unexpected input with resource implications is treated as signal and encoded; everything else is noise and safely ignored.
Note what this does to the name. On Craig’s reading the network detects salience — salient things arrive and it notices them. On Barrett’s the network confers salience: nothing is intrinsically salient, and what counts as signal is decided by a prediction about what will matter to the body budget. The construct’s name encodes the reading Barrett rejects.
A second, less-cited route she proposes: limbic regions in the network may signal precision indirectly, by modulating the thalamic reticular nucleus that encircles the thalamus and gates what reaches cortex at all (Zikopoulos & Barbas, 2006, 2012; John et al., 2016). If that holds, precision-weighting is happening before the thalamocortical relay — earlier than any of the wiki’s other predictive sources place it.
Three readings of one network
| source | the salience network… |
|---|---|
| Craig (2009) | detects — ascending salient signals converge on the AIC/ACC dyad |
| Seth (2013) | compares — predictions are generated, compared against input, updated |
| Barrett (2017) | assigns — sets the gain that decides which errors count, before they arrive |
The first two are the poles of feedforward-vs-predictive-interoception. Barrett’s is on Seth’s side of that dispute but is not the same claim: comparison is what a comparator does with an error signal; precision is what determines whether the error signal is worth computing. She and Seth agree on the architecture and divide the labour differently within it — see the precision material on predictive-coding and seth-friston-2016-active-interoceptive-inference, where precision does the homeostasis/allostasis work instead.
Not an emotion network — a body-budgeting network
Barrett’s Figure 6 presents the salience network and the default mode network together as a single large-scale system for allostasis and interoception, interconnected by rich-club hubs (ventral anterior insula, midcingulate, parahippocampal gyrus, postcentral gyrus, cuneus) and reaching subcortically to PAG, PBN, NTS, ventral striatum, hypothalamus and thalamus — mapped in 280 participants, replicated in 270 (anatomy credited to Kleckner et al., “in press” at the time).
This reframes the convergence recorded in the section above. The wiki notes that three independent framings — functional-network (Seeley), meta-analytic (core affect), cytoarchitectural (VMAs) — land on nearly the same regions, and treats that as a striking coincidence between literatures. Barrett’s answer is that it is not a coincidence and not about emotion: all three were measuring the system that budgets the body, and calling it emotion because that is what the regions have traditionally been assumed to do. See allostasis, theory-of-constructed-emotion.
She adds the corresponding disclaimer, which the wiki should keep: this does not mean the default mode and salience networks are allostasis networks and nothing else. “These (and other) domain-general networks can be mapped to many psychological categories at the same time.” Domain-generality is the point of them, and a network that is for one thing would be exactly the sort of claim her theory denies.
The network at work: compassion, empathy, and salience to another (Lutz et al. 2008)
The wiki’s other pages treat the network detecting/assigning salience in one’s own homeostatic interest. Lutz et al. (2008) shows the same anchors — AIC, ACC, amygdala — recruited by salience in another person’s state, and trained up by compassion meditation. Generating the compassionate state up-regulated AIC/ACC to emotional human vocalizations (especially distress in experts), and the broader expertise effect extended the network’s classic attentional-detection arm — right TPJ and right IFG, “an attentional system specialized to detect behaviourally relevant stimuli” (Corbetta & Shulman 2002), plus bilateral amygdalae for appraisal of emotional stimuli. Experts flipped from a negative response to the sounds at rest to a positive one during meditation, which the authors read as being more primed to detect the salient suffering of others. So compassion training reads as tuning the salience network to treat another’s distress as behaviourally relevant — empathy as trained salience-detection, with the interoceptive AIC as its affective core.
A fourth reading, and the one that gives the network a clinical job (Weng et al. 2021)
Weng et al. (2021) introduce the network in Menon & Uddin’s (2010) switching form, which the wiki had only met glancingly through Craig’s remark about the right AIC. Four functions in that model: bottom-up detection of salient events; switching between large-scale networks to bring attention and working memory online when one is detected; anterior–posterior insula interaction modulating autonomic reactivity to salient stimuli; and strong ACC coupling for rapid access to the motor system.
What Weng et al. add is a use for it. Their claim is that the salience network is the bridge between interoception and executive function — the reason attentional control, studied almost entirely with visual and auditory stimuli (Corbetta & Shulman 2002; Petersen & Posner 2012), can be turned on internal signals at all. Psychological interventions such as meditation are then read as using executive-control networks to increase engagement of the interoception and salience networks, which share the anterior insula as a hub, while disengaging the DMN.
So the wiki now has four readings:
| source | the salience network… |
|---|---|
| Craig (2009) | detects |
| Seth (2013) | compares |
| Barrett (2017) | assigns |
| Weng et al. (2021) | switches — and is therefore trainable |
The fourth is the least theoretically ambitious and the most practically consequential: if the network’s job is switching, then an intervention that repeatedly practises switching attention to interoceptive signals has an obvious target, which is the mechanistic bet behind every contemplative entry in the applications folder. Held as a rationale rather than a demonstration — nothing in the review shows the switching function itself changing with training.
It is two networks, not one, and only half of it is homologous
A structural correction from Chen et al. (2021) that bears on every reading above. The insula is described as part of the ventral salience network, for arousal-based affective experience — and that ventral network “is largely homologous between monkeys and humans,” whereas the dorsal network for attentional control is much more developed in humans.
Everything on this page has been written about “the salience network” as a single object. If the affective/arousal division and the attentional-control division have different evolutionary depths, three of the four readings above are affected in different ways:
- The switching account (Weng et al., Menon & Uddin) is largely a claim about attentional control — the more human-specific half — which makes it harder to model in animals but does not weaken it.
- The body-budgeting account (Barrett) is largely a claim about the ventral, homologous half, which is the half rodent and monkey work can speak to.
- Any inference that runs from a rodent salience-network finding to a human attentional claim is crossing the seam.
The wiki should stop treating “salience network” as one term when the source is doing cross-species work. Related species cautions are itemized at insular-cortex.
The network in its failure mode: frontotemporal dementia (Farb et al. 2012)
Every reading above describes the network working. Farb et al. (2012) show what its dissolution does, in a clinical population where the salience network degenerates. In both frontotemporal-dementia variants (bvFTD and semantic dementia), resting-state analysis found reduced frontolimbic connectivity — the salience-network core (anterior insula, ACC, thalamus, striatum) losing its coherence — alongside a compensatory-looking prefrontal hyperconnectivity. The right anterior insula specifically showed reduced low-frequency power (fALFF), reduced local coherence (REHO) and severed PFC connectivity: the apex of the salience hierarchy going offline.
Two things this adds to the page. First, a behavioural stake: prefrontal hyperconnectivity tracked apathy, while frontolimbic disconnection tracked lower disinhibition — the disinhibited symptoms need an intact salience/limbic channel to be produced at all. Second, the salience network’s proposed switching role (Menon & Uddin, above) is here observed failing: as the network dissolves it stops regulating the DMN, and bvFTD showed unchecked DMN elevation (right angular gyrus) tracking stereotypy — automaticity freed from the salience network’s veto. So the fourth (switching) reading gains its clinical converse: when switching breaks, habit and self-narrative run without the salience network to interrupt them.
Frontal salience self-referential activity marks relapse, but its coupling does not (Wu et al. 2026)
Wu et al. (2026) put the salience network’s frontal node (ACC) to a preregistered clinical test alongside the frontal DMN. Dysphoric self-referential activity (Self-Negative minus Self-Positive) in the frontal salience ROI was associated with relapse status (b=0.36 [0.09, 0.63]), and relapsers showed higher baseline salience-network dysphoric activity. But — as with the DMN — frontal SN activity did not independently predict relapse timing in survival models, no Relapse×Time interaction reached significance for it (unlike the DMN), and crucially frontal DMN–SN connectivity, and within-SN connectivity, were unrelated to relapse.
This is a negative result against the dorsal-nexus account (Sheline et al. 2010; Bertocci et al. 2023), on which depression vulnerability is carried by heightened DMN↔salience coupling. In this remitted cohort the coupling was inert; the authors read prefrontal activation during self-referential processing as more relapse-relevant than inter-network connectivity per se. It bears on the switching/precision readings above: the salience network’s activation flags a relapser, but neither its internal coherence nor its coupling to the DMN forecast the episode here — a boundary on how far the “salience↔DMN antagonism” story (Barrett; Weng) extends into prospective clinical prediction.
The network assigning salience to the outside — sensory over-responsivity (Schwarzlose et al. 2023)
Every reading above is about the network detecting or assigning salience to internal or emotional signals. Schwarzlose et al. (2023) show the same nodes implicated when the over-attended signal is exteroceptive — and it is the wiki’s cleanest case of Barrett’s “assigns” reading operating on external input. In 11,210 ABCD children, sensory over-responsivity (distress at innocuous sounds, textures, smells) came with enhanced sensorimotor-hand↔salience resting-state connectivity and enhanced within–ventral-attention connectivity, plus enhanced cingulo-opercular↔amygdala coupling. The authors read this exactly as this page’s Barrett section would predict: the salience/attention system directing attention toward sensory input that should have been treated as noise and ignored — over-precision on exteroceptive prediction error.
Two things this adds. First, it is a developmental, clinical, large-N instance of the “assigns” job (versus the healthy-adult and lesion cases above). Second, it sharpens a caution that runs through interoceptive-psychopathology: if the same salience machinery mis-assigns relevance to exteroceptive input in one transdiagnostic sensory syndrome and to interoceptive input in another (anxiety-sensitivity, panic), then salience-precision failure may be a domain-general vulnerability rather than an interoception-specific one — see interoception-exteroception-boundary, is-interoception-domain-general. Held loosely: the SOR measure is a single parent-report item, and the FC is correlational.
Note
The salience network is a widely cited construct that will recur in future ingests (anxiety, addiction, psychosis all implicate it); this page is intentionally concise and will grow as sources accumulate.