Xinyi Yuan¹, Sabyasachi Shivkumar², Maeve Silverman¹, Ralf M. Haefner¹; ¹University of Rochester, ²Columbia University

Retinal motion is a key cue for inferring object motion, scene parsing, and object relationships. Recent work has modeled this as hierarchical causal inference (Shivkumar et al., 2024), assuming the brain infers causal links between moving elements in a scene in order to construct the reference frame within which to represent motion. For example, a person’s motion should be interpreted in the frame of a moving escalator if they are walking on it, but not otherwise. We tested this by manipulating the causal structure of a scene using a center-surround motion stimulus. Typically, perceived center motion is biased towards the surround's direction for small relative differences (integration) and away for larger differences (segmentation). If causal inference drives this bias, it should weaken when evidence against a causal link is provided and strengthen when evidence supports it. Four human observers reported their perceived motion direction using a dial under four conditions: (1) both center and surround reversed direction in a way consistent with a causal link, (2) the center did not reverse while the surround did, weakening causal evidence, and (3)&(4) observers controlled the reversal point using a button press, further modulating causal evidence. Results showed biases that clearly differed between causal conditions. Interestingly, direct intervention by observers provided little additional causal information compared to the observation of common motion alone. Notably, we found substantial inter-observer variability: two observers showed a reduction in surround influence in conditions 2&4 compared to 1&3, reflecting the differences in causal information, while the effect was less clear for the other two. Furthermore, even in the acausal conditions 2&4 a bias remained. These findings highlight unexpected variability in how observers interpret common motion and suggest that retinal motion provides robust causal information, with limited additional benefit from direct stimulus control.

Acknowledgements: We acknowledge funding support from NIH/U19 NS118246, and NSF/CAREER IIS-2143440.