Fig. 3: The model quantitatively explains human perception of nested and ambiguous motion scenes.

a Stochastic motion stimulus from Yang et al.¹⁷ consisting of three dots rotating on a circle. b Each trial followed one of four motion structures. If clustered motion was present (C or H structure), any pair of dots could form the cluster. c Confusion matrix of human responses, averaged over all 12 participants. d Models for predicting human responses. Yang et al. employed a Bayesian ideal observer as the basis for fitting a participant-specific choice model. Our model, in contrast, calculates the likelihood for each structure from the motion strengths, λ_t, at trial end and then fits the same choice model as Yang et al. for translating probabilities into human responses. e Confusion matrix of our model. f Log-likelihood of human responses relative to chance level, for both models. The analyses in panels e and f are leave-one-out cross-validated to prevent overfitting. Source data are provided as a Source Data file.