Individual differences in learning predict the return of fear

Samuel J. Gershman, Catherine A. HartleyLearning 2015 · DOI

This paper, published in Learning & Behavior, offers a computational perspective on a classic psychological puzzle: why does fear often return after it has seemingly been "extinguished" through therapy or training? By combining Pavlovian conditioning experiments with Bayesian modeling, Gershman and Hartley demonstrate that the return of fear is not random, but a predictable result of how different individuals structurally organize their memories.

TL;DR

The return of fear (spontaneous recovery) depends on whether an individual "unlearns" a fear memory or simply creates a new, competing "safety" memory. Using a latent cause model, the authors show that "two-state" learners who categorize extinction as a new experience are significantly more likely to see their fear return than "one-state" learners who update their original fear memory.

What Problem Does This Paper Address?

In clinical settings, anxiety disorders are often treated with extinction-based therapies (repeatedly exposing a patient to a feared stimulus without the negative outcome). While often successful in the short term, many patients experience spontaneous recovery—the return of fear after time has passed.

The field has long debated whether extinction represents "unlearning" the original fear or "new learning" of safety that merely inhibits the old fear. This paper addresses the gap in understanding individual differences: why do some people seem to "unlearn" (resilient) while others only "inhibit" (vulnerable to relapse)?

Methods

The authors re-analyzed a dataset of 106 human participants who underwent a two-day fear conditioning paradigm:

  1. Acquisition (Day 1): A neutral stimulus (CSa) was paired with a mild electric shock.
  2. Extinction (Day 1): CSa was presented repeatedly without the shock.
  3. Recall (Day 2): Participants returned to test for spontaneous recovery.

The Innovation: The authors applied a Latent Cause Model to the participants' Skin Conductance Responses (SCR). Unlike traditional models (like Rescorla-Wagner) that assume a single associative strength, this Bayesian model posits that learners try to infer the "hidden state" of the world.

  • If a learner perceives a large enough "prediction error" (e.g., the shock suddenly stops), the model decides whether to update the current state (unlearning) or create a brand-new latent cause (new safety memory).
  • A key parameter, $\alpha$ (alpha), represents an individual's prior belief in the complexity of the environment. A high $\alpha$ leads to "state splitting" (creating new memories), while a low $\alpha$ leads to "state merging" (updating old ones).

Datasets

The study utilized an existing dataset from Hartley et al. (2012), consisting of 106 healthy human volunteers (ages 18–35). Fear was measured via differential SCR (the response to the threat stimulus minus the response to a safe control stimulus).

Key Insights and Ideas

  • Structural Inference as a Predictor: The return of fear is a byproduct of how we cluster experiences. If you categorize "extinction" as a separate "chapter" from "acquisition," the original fear chapter remains intact and can be retrieved later.
  • The "One-State" vs. "Two-State" Split: The model identified two distinct groups. Most participants (83%) were "one-state" learners who merged acquisition and extinction. A smaller group (17%) were "two-state" learners who segmented them.
  • Computational vs. Clinical Markers: Interestingly, the model’s internal parameters (specifically the "log Bayes factor," which measures the preference for multiple states) were better predictors of fear recovery than traditional measures like trait anxiety, age, sex, or specific genetic markers (STPP G alleles).

Key Results

  • Behavioral Divergence: The "two-state" group showed significant spontaneous recovery on Day 2 ($p < .0001$), whereas the "one-state" group showed essentially none ($p = .79$).
  • Model Fit: The latent cause model outperformed the classic Rescorla-Wagner model for every single participant. The improvement was particularly dramatic for the "two-state" group ($t(104) = 8.79, p < .00001$).
  • Correlation: There was a significant positive correlation ($r = .40, p < .0001$) between a participant's model-derived preference for multiple states and the actual amount of fear that returned the next day.

Limitations and Future Work

  • Healthy vs. Clinical Populations: The study used healthy volunteers. It remains to be seen if individuals with diagnosed PTSD or anxiety disorders show a higher baseline $\alpha$ (a stronger tendency to split states), which would explain their clinical vulnerability.
  • Neural Mechanisms: While the model suggests "states," the paper does not identify the specific brain regions responsible. The authors speculate that the hippocampus (involved in pattern separation) and the amygdala/mPFC circuit are likely candidates.
  • Speculative Nature of "Unlearning": The authors acknowledge that "unlearning" is a strong claim. Other assays (like "reinstatement" or "renewal") might still reveal hidden fear memories even in the "one-state" group.

This work suggests that the future of anxiety treatment might lie in "nudging" the brain's clustering mechanisms—forcing the learner to update the old fear memory rather than creating a new, fragile safety memory.

Summary updated

BibTeX
@inproceedings{gershman2015individual,
  author = {Samuel J. Gershman and Catherine A. Hartley},
  title = {{Individual differences in learning predict the return of fear}},
  booktitle = {Learning},
  year = {2015},
  doi = {10.3758/s13420-015-0176-z},
}

Loading related content...

Figure