Researcher(s)
- Sarah Rowe, Psychology, University of Delaware
Faculty Mentor(s)
- Timothy Vickery, Psychology, University of Delaware
Abstract
Visual statistical learning (VSL) is the brain’s ability to automatically and unintentionally learn patterns, sequences, and statistical regularities in the visual environment. Previous research shows that shared category membership strengthens VSL for items that are associated with one another. However, little research has examined whether this is due to category membership itself or to individuals perceiving stimuli from the same natural category as more visually similar. This study tests whether VSL is affected by the degree of similarity between scenes. Each session included two phases: a familiarization phase (one-back task) and a test phase (forced-choice recognition task). During the one-back task, participants viewed a stream of artificial scenes generated by generative adversarial networks (GANs). Scene similarity was manipulated by varying the distance between images in a 512-dimensional latent space, with similarity ratings validated with a separate sample of naïve participants. Unknown to participants, similar and dissimilar images were presented in pairs, with some images always preceding the same image in the stream. They were shown a mix of artificially generated landscape and interior scenes. For this task, they were occasionally asked to indicate whether the scene currently on screen was the same as the scene shown immediately prior. After completing that task, participants were informed of the hidden pairings and performed a forced-choice recognition task, identifying the true pairs previously shown from a foil pair of two images from different pairs, based on which pair felt more familiar. Results showed no statistically significant effect of similarity on pair selection accuracy (P = .727). However, overall recognition rates were very low (not significantly above chance) for every pair type. These results suggest poor statistical learning for this type of stimulus compared with typical stimulus sets used in VSL studies. Future improvements suggested by our results include increasing stimulus exposure during training, altering the cover task, and testing a larger sample.