Abstract

According to the Theory of Constructed Emotion, visual emotion codes in the brain are regularized by abstract emotion concepts from memory. While neuroimaging studies show that similar concepts elicit similar neural patterns—suggestive of predictive coding through error minimization—it remains unclear whether these codes are also shaped by developmental experience. Such experience-dependent specialization could enable efficient emotion recognition without constant memory system engagement, yet this mechanism is rarely investigated. Here, we introduce a deep neural network (DNN) framework to probe this process. First, we trained a concept-regularized DNN to model the brain’s visual emotion codes. Next, we perturbed its regularization strength to simulate impaired emotion processing in depression. From this mechanism, we derived a computational phenotype and validated its utility in population-based and clinical cohorts. Our findings illuminate how concept regularization underpins facial emotion recognition and its dysfunction in depression, offering a novel DNN -based lens for computational psychiatry.