Abstract
<title>Abstract</title> <p> Predicting false alarms, i.e., when novel images are mistakenly identified as familiar, is crucial for understanding visual memory bias. Prior studies show that some images consistently elicit more false alarms than others, so to model these patterns, we introduce <italic>FalseResMem</italic> , the first neural network designed to predict image-level false alarm rate (FAR). The model combines ImageNet-pretrained ResNet50 features with a retrained AlexNet-like architecture to capture image properties that drive false recognitions. Trained on a large-scale object memory dataset, FalseResMem achieves consistent performance (Spearman’s rank correlation = 0.48 ± 0.13; max = 0.58, 10-fold cross-validation) and FalseResMem’s predictions successfully generalize to other image categories, including scenes, art, faces, and symbols, demonstrating its robustness. This contribution highlights the potential of integrating error-based metrics into visual memory modeling through a simple FalseResMem package, offering researchers a practical way to predict false alarms and investigate the image features that systematically shape false memories. </p>