Model-Free and Model-Based Learning in Human Fear Conditioning.

Abstract

Learning to predict threat and to adapt behavior accordingly is critical to human survival. This type of learning is governed by distinct systems in the brain. Model-free learning is reflexive and computationally efficient, allowing for rapid responses based on past experience, but is also resistant to environmental change. In contrast, model-based learning is more flexible and relies on an internal model of the environment that can be updated without direct experience. To enhance the translational value of human fear-conditioning research and facilitate the interpretation of fear-reduction interventions, we examined whether commonly used measures of conditioned responding reflect model-free and/or model-based learning. Based on prior observations and the defensive-reflex nature of the fear-potentiated startle (FPS), we hypothesized that FPS responses would reflect model-free learning, whereas skin conductance responses (SCRs) would reflect model-based learning. To test this hypothesis, we decoupled the experienced US value (i.e., model-free) from the expected value (i.e., model-based) by US devaluation/revaluation manipulations delivered via instructions (Exp1), physical disconnection of the US electrode (Exp2), or their combination (Exp3). By design, instantaneous changes in conditioned responding, prior to any new learning by experience, must indicate model-based updating rather than model-free cached value. Across three fear-conditioning experiments (total n = 131), we found that SCRs most likely reflected model-based learning, whereas FPS responses reflected aspects of both model-free and model-based learning. These results show that common indices of conditioned fear are not simply interchangeable: they tap distinct learning systems, and treating them as equivalent risks missing the more affective, experience-driven components of fear-thereby limiting translation to clinical intervention.