Object motion amplitude and generalization in contextual cueing

Our environment contains statistical regularities. Objects often appear repeatedly in similar locations or under similar situations. People can learn these regularities without intention (i.e., statistical learning). This learning can shape how attention is allocated and can therefore influence behavior. Some studies suggest that statistical learning can generalize beyond what was directly experienced to similar new instances (Chun & Jiang, 1998). At the same time, other studies report cases where generalization is weak or absent (Chua & Chun, 2003; Endo & Takeda, 2005; Makovski & Jiang, 2010; Zellin et al., 2013). These mixed findings raise the question of what factors shape the extent of generalization in statistical learning. Our previous study suggested that variability across trials is an important factor shaping generalization (Higuchi et al., 2020). When people are repeatedly exposed to instances that vary slightly from trial to trial (i.e., slightly different spatial configurations), the brain can learn statistical structure that incorporates this variability, supporting generalization to new but similar instances. In contrast, when the same instance is repeated across trials with little variability (i.e., an identical spatial configuration), learning tends to become tied to that specific instance, and generalization is reduced. In our previous study, “variability” referred to small differences in static spatial configurations across repeated exposures (i.e., across trials in the lab). In real-world environments, however, objects often move, and their spatial relations can change over time even within a single scene. In dynamic environments, variability therefore arises not only across scenes but also within a scene as changes unfold over time. In the present study, we manipulate this within-scene variability as motion amplitude (i.e., the range of movement an object can take). We test the following hypothesis: in scenes composed of moving objects, the brain learns statistical structure that incorporates motion amplitude, and the generalization of statistical learning depends on the magnitude of motion amplitude. We examine this hypothesis using a statistical learning paradigm based on spatial configurations—specifically, a contextual cueing paradigm. In repeated configurations, we manipulate how far each object can move within its placeholder (i.e., motion amplitude). With large motion amplitude, the spatial relations among objects change substantially over time, so observers are exposed not to a single identical configuration but to a set of related configurations that vary within a bounded range. Under this condition, learning is expected to generalize more easily to new but similar configurations that share the same statistical structure. With small motion amplitude, within-configuration variability is limited, learning tends to become more tied to specific configurations, and generalization is expected to be weaker.