Distinct representational properties of cues and contexts shape fear and reversal learning
Hypotheses The paper tests whether the fear network encodes threatening cues through a generalized between-item representation whenever threat is acquired, whether reversal recruits two simultaneous representational strategies — generalization for currently dangerous cues and item-specific stabilization for cues whose valence changes — in dissociable regions, and whether prefrontal context-specific coding is the substrate linking context distinctiveness to context-dependent fear renewal.
Claims RSA on fMRI data across acquisition, reversal, and test phases establishes CS+ > CS- cue generalization in dACC, SFG, caudate, and insula during acquisition; during reversal, newly dangerous CS-+ cues acquire a generalized fear-network pattern while changing-contingency cues exhibit elevated item-specific stability in precuneus and IFG. Context representations become more distinct in dmPFC and lateral PFC during reversal, and PFC context specificity predicts subsequent reinstatement of acquisition fear traces at test.
Inferences The brain adapts to shifting contingencies through a flexible composition of generalization and item-specific coding rather than a single representational scheme, and prefrontal context coding is implicated as a candidate neural mechanism of fear renewal — a framing with direct implications for why exposure therapy gains do not always transfer across contexts.
▸ Summary
▸Hypotheses tested
PFC context-specific coding is the neural substrate for context-dependent fear renewal — participants with more distinct PFC context representations should show stronger reinstatement of acquisition fear memory traces.
The renewal hypothesis predicts that PFC context representations should become more distinct during reversal than during acquisition, when context becomes behaviorally relevant.
Tested by
Neural representations of contexts become more distinct (context-specific) during reversal learning compared to acquisition, particularly in prefrontal cortex, reflecting the need to separate safe and dangerous environments.
The renewal hypothesis predicts that PFC context specificity during reversal should covary across subjects with subsequent reinstatement of acquisition fear memory traces at test.
Tested by
Prefrontal cortex context-specificity during reversal learning predicts subsequent fear renewal at test, linking context-specific neural coding to behavioral expression of fear.
Reversal learning recruits two simultaneous strategies — generalization for currently dangerous cues and item-specific updating for changing-valence cues — in dissociable brain regions.
The generalization hypothesis predicts that during reversal, the newly dangerous CS-+ cue should acquire generalized fear-network representations like CS++ during acquisition.
Tested by
During reversal, the newly dangerous cue (CS-+) acquires generalized neural representations similar to the originally feared CS++, mirroring the initial acquisition pattern in fear network regions.
The dual-strategy hypothesis predicts that changing-valence cues should show elevated item-specific representations during reversal in precuneus, IFG, and PFC.
Tested by
During reversal learning, item-specific (stable across phases) representations emerge in precuneus and prefrontal cortex for cues that change their threat value (CS+-), distinguishing the changing cue from the stable threats.
Fear learning encodes threatening cues with shared, generalized representations in the fear network — the same generalization mechanism is recruited whenever a cue acquires threat value, in both initial acquisition and reversal.
The generalization hypothesis predicts that during acquisition, CS+ items should show greater between-item pattern similarity than CS- items in fear-network regions (dACC, SFG, caudate, insula).
Tested by
Neural representations of threatening cues (CS++) generalize across items during fear acquisition: RSA shows that patterns for different CS++ items become more similar to each other in the fear network, indicating a shared threat-cue representation.
The generalization hypothesis predicts that during reversal, the newly dangerous CS-+ cue should acquire generalized fear-network representations like CS++ during acquisition.
Tested by
During reversal, the newly dangerous cue (CS-+) acquires generalized neural representations similar to the originally feared CS++, mirroring the initial acquisition pattern in fear network regions.
▸Dissociations
In dmPFC during test_old, generalized reinstatement is higher for acquisition memory traces than for test_new memory traces (F(2,259)=4.01, p<0.05; t(259)=2.96, p<0.05), showing that dmPFC preferentially reinstates category-level (generalized) acquisition representations.
In IFG during test_old, item reinstatement is higher for reversal memory traces than for acquisition or test_new memory traces (F(2,253)=5.50, p<0.01), showing that IFG preferentially reinstates the most recently learned item-specific representations.
During reversal learning, item-specific (stable across phases) representations emerge in precuneus and prefrontal cortex for cues that change their threat value (CS+-), distinguishing the changing cue from the stable threats.
During reversal, the newly dangerous cue (CS-+) acquires generalized neural representations similar to the originally feared CS++, mirroring the initial acquisition pattern in fear network regions.
During fear acquisition, item stability (within-cue neural pattern similarity) does not differ between CS+ and CS- cues anywhere in the brain, showing that threat learning selectively increases cross-item generalization but not single-item representational consistency.
Neural representations of threatening cues (CS++) generalize across items during fear acquisition: RSA shows that patterns for different CS++ items become more similar to each other in the fear network, indicating a shared threat-cue representation.
▸Interpretations
During reversal, cues that were threatening during acquisition but not currently threatening (CS++) vs those that were safe during acquisition and currently threatening (CS-+) still show fear network activation — (CS++ > CS+-) > (CS-+ > CS--) — though to a lesser extent than the current-threat contrast, suggesting a lingering prior fear memory trace.
▸Synthesis claims
Reversal learning recruits two distinct representational strategies simultaneously: generalization (treating newly dangerous CS-+ like old CS++) for currently dangerous cues, and item-specific updating (distinguishing CS+- from CS++) for cues with changing threat value.
▸Standalone empirical findings
US expectancy ratings follow the hierarchy CS++ > CS+- > CS-+ > CS-- across all experimental phases (LME: CS type F=479.35, p<0.0001; phase F=125.6, p<0.001; interaction p<0.001), confirming participants learned threat contingencies and their reversals.
Higher reversal context specificity (PFC) interacts with CS type to predict generalized reinstatement of acquisition memory traces in ACC/SFG (favoring initially threatening CS+-, t(22)=6.25, p<0.05) and precuneus (favoring initially safe CS-+, t(22)=4.89, p<0.01), with opposite directions across regions.
Higher reversal context specificity (PFC) predicts greater item reinstatement of CS-+ than CS+- reversal memory traces in dmPFC during test_old (t(22)=5.56, p<0.05), favoring reinstatement of threatening reversal memories in the same region that generalizes acquisition traces.
During fear acquisition, threatening cues (CS+) produce significantly greater BOLD activation than safe cues (CS-) in dACC, superior frontal gyrus, caudate nucleus, and middle temporal gyrus, replicating the canonical fear network activation pattern.
During reversal, cue generalization for CS++ (always threatening) vs CS-- (always safe) is elevated only in dACC, not in the broader fear network regions where this effect was present during acquisition, suggesting that consistent threat representations are maintained more narrowly.
During reversal, currently threatening cues (CS++ and CS-+) produce greater BOLD activation than non-threatening cues (CS+- and CS--) across the same fear network regions as acquisition (dACC, SFG, MTG, IFG), reflecting rapid updating of neural threat responses.
Item stability (but not cue generalization) persists into test phases in the absence of a US: CS+- > CS++ item stability in MTG at test_new, and CS++ > CS-- item stability in inferior temporal gyrus at test_old, showing that individual-item memory traces outlast the training context.
▸Methodological warrants
All RSA pattern similarity analyses use only unreinforced trials (no US delivered), excluding reinforced trials to avoid US-driven BOLD confounds; this exclusion reduces the effective trial count and may selectively remove trials with strongest fear responses.
ROI-based RSA analyses (Fig 4B, Fig 5D) use ROIs derived from statistically significant clusters in the preceding searchlight analyses on the same dataset, without an independent ROI definition, creating a circularity that inflates the expected significance of ROI tests.
▸All claims (alphabetical)
- behavioral-learning-confirms-contingencies fig2A
- context-specificity-increases-reversal fig5B
- context-specificity-predicts-acquisition-reinstatement-regional-dissociation fig5Di
- context-specificity-predicts-reinstatement-new-context-mtg fig5Diii
- context-specificity-predicts-reversal-reinstatement-dmpfc fig5Dii
- cs-plus-univariate-fear-network-acquisition fig2Bi
- cue-generalization-increases-acquisition fig3A
- cue-generalization-limited-dacc-reversal-consistent fig3Bi
- current-threat-activates-fear-network-reversal fig2Bii
- dmpfc-reinstates-acquisition-traces-generalized fig4Bii
- dual-strategy-reversal-generalize-plus-specify fig4 (synthesis)
- generalized-pattern-cs-minus-plus-reversal fig3Bii
- hypothesis-context-specificity-supports-renewal hypothesis
- hypothesis-dual-strategy-reversal hypothesis
- hypothesis-fear-network-generalizes-threat-cues hypothesis
- ifg-reinstates-reversal-traces-item-specific fig4Bi
- item-stability-persists-test-phases fig3C, fig3D
- item-stability-precuneus-pfc-reversal fig4
- lss-unreinforced-trials-only methods (RSA section)
- no-bold-differences-test-phases fig2B (test phases)
- no-item-stability-difference-acquisition fig3A
- pfc-context-specificity-predicts-renewal fig5
- prediction-context-specificity-increases-during-reversal prediction
- prediction-cs-minus-plus-generalizes-reversal prediction
- prediction-cue-generalization-fear-network-acquisition prediction
- prediction-item-stability-changing-cues-reversal prediction
- prediction-pfc-context-specificity-tracks-renewal prediction
- prior-threat-activates-fear-network-weakly fig2Biii
- rsa-roi-derived-from-searchlight fig4A (ROI definition procedure)
- trace-conditioning-hippocampus-engaged fig1 (methods)
Abstract mapped to claims
The paper's abstract is shown with each sentence linked to the claim(s) it represents in the dependency graph. Hover or click a sentence to highlight the corresponding claim cards. Below: what the graph contains that the abstract leaves out, and vice versa.
1When we learn that something is dangerous, a fear memory is formed. 2However, this memory is not fixed and can be updated through new experiences, such as learning that the threat is no longer present. 3This process of updating, known as extinction or reversal learning, is highly dependent on the context in which it occurs. 4How the brain represents cues, contexts, and their changing threat value remains a major question. 5Here, we used functional magnetic resonance imaging and a novel fear learning paradigm to track the neural representations of stimuli across fear acquisition, reversal, and test phases. 6We found that initial fear learning creates generalized neural representations for all threatening cues in the brain’s fear network. 7During reversal learning, when threat contingencies switched for some of the cues, two distinct representational strategies were observed. 8On the one hand, we still identified generalized patterns for currently threatening cues, whereas on the other hand, we observed highly stable representations of individual cues (i.e. item-specific) that changed their valence, particularly in the precuneus and prefrontal cortex. 9Furthermore, we observed that the brain represents contexts more distinctly during reversal learning. 10Furthermore, additional exploratory analyses showed that the degree of this context specificity in the prefrontal cortex predicted the subsequent return of fear, providing a potential neural mechanism for fear renewal. 11Our findings reveal that the brain uses a flexible combination of generalized and specific representations to adapt to a changing world, shedding new light on the mechanisms that support cognitive flexibility and the treatment of anxiety disorders via exposure therapy.
- E1 behavioral-learning-confirms-contingencies fig2A US expectancy ratings follow the hierarchy CS++ > CS+- > CS-+ > CS-- across all experimental phases (LME: CS type F=479.35, p<0.0001; phase F=125.6, p<0.001; interaction p<0.001), confirming participants learned threat contingencies and their reversals.
- E4 cs-plus-univariate-fear-network-acquisition fig2Bi During fear acquisition, threatening cues (CS+) produce significantly greater BOLD activation than safe cues (CS-) in dACC, superior frontal gyrus, caudate nucleus, and middle temporal gyrus, replicating the canonical fear network activation pattern.
- E6 current-threat-activates-fear-network-reversal fig2Bii During reversal, currently threatening cues (CS++ and CS-+) produce greater BOLD activation than non-threatening cues (CS+- and CS--) across the same fear network regions as acquisition (dACC, SFG, MTG, IFG), reflecting rapid updating of neural threat responses.
- I1 prior-threat-activates-fear-network-weakly fig2Biii During reversal, cues that were threatening during acquisition but not currently threatening (CS++) vs those that were safe during acquisition and currently threatening (CS-+) still show fear network activation — (CS++ > CS+-) > (CS-+ > CS--) — though to a lesser extent than the current-threat contrast, suggesting a lingering prior fear memory trace.
- C2 no-bold-differences-test-phases fig2B (test phases) During both test phases (test_new and test_old), no significant BOLD activation differences between any CS type contrasts are found, despite significant US expectancy differences at the behavioral level, motivating RSA over univariate analysis.
- D3 no-item-stability-difference-acquisition fig3A During fear acquisition, item stability (within-cue neural pattern similarity) does not differ between CS+ and CS- cues anywhere in the brain, showing that threat learning selectively increases cross-item generalization but not single-item representational consistency.
- E5 cue-generalization-limited-dacc-reversal-consistent fig3Bi During reversal, cue generalization for CS++ (always threatening) vs CS-- (always safe) is elevated only in dACC, not in the broader fear network regions where this effect was present during acquisition, suggesting that consistent threat representations are maintained more narrowly.
- E7 item-stability-persists-test-phases fig3C, fig3D Item stability (but not cue generalization) persists into test phases in the absence of a US: CS+- > CS++ item stability in MTG at test_new, and CS++ > CS-- item stability in inferior temporal gyrus at test_old, showing that individual-item memory traces outlast the training context.
- D1 ifg-reinstates-reversal-traces-item-specific fig4Bi In IFG during test_old, item reinstatement is higher for reversal memory traces than for acquisition or test_new memory traces (F(2,253)=5.50, p<0.01), showing that IFG preferentially reinstates the most recently learned item-specific representations.
- D1 dmpfc-reinstates-acquisition-traces-generalized fig4Bii In dmPFC during test_old, generalized reinstatement is higher for acquisition memory traces than for test_new memory traces (F(2,259)=4.01, p<0.05; t(259)=2.96, p<0.05), showing that dmPFC preferentially reinstates category-level (generalized) acquisition representations.
- E2 context-specificity-predicts-acquisition-reinstatement-regional-dissociation fig5Di Higher reversal context specificity (PFC) interacts with CS type to predict generalized reinstatement of acquisition memory traces in ACC/SFG (favoring initially threatening CS+-, t(22)=6.25, p<0.05) and precuneus (favoring initially safe CS-+, t(22)=4.89, p<0.01), with opposite directions across regions.
- E3 context-specificity-predicts-reversal-reinstatement-dmpfc fig5Dii Higher reversal context specificity (PFC) predicts greater item reinstatement of CS-+ than CS+- reversal memory traces in dmPFC during test_old (t(22)=5.56, p<0.05), favoring reinstatement of threatening reversal memories in the same region that generalizes acquisition traces.
- C1 context-specificity-predicts-reinstatement-new-context-mtg fig5Diii Higher reversal context specificity (PFC) predicts greater item reinstatement of CS-+ than CS+- acquisition memory traces in MTG during test_new (t(22)=2.51, p<0.05), the phase with entirely new contexts, suggesting context specificity generalizes beyond the training context.