Abstract
Introduction: The Bayesian ideal searcher (IS; Najemnik & Geisler, 2005) is an essential normative model to understand how humans make eye movements during search. The IS uses knowledge of variations of target detectability with retinal eccentricity, and calculates fixation locations that maximize target detection accuracy. The IS formulation assumes that the internal responses are independent across fixations which occurs with temporally changing image noise or dominant dynamic internal noise. When searching static images, human responses across fixations are often partially correlated. In addition, image-computable foveated search models (Akbas & Eckstein, 2017, Lago el al., 2020) result in response correlations across fixations that depend on the image spatial statistics and the saccade amplitudes. What are the optimal eye movements with inter-saccade correlations? We developed an IS that accounted for the inter-saccade correlations (IS-COR) and compared it to the standard IS. Method: We used an image-computable foveated search model and the covariance between template responses over fixations was used to predict the optimal fixations of the IS-COR. We varied the noise power spectra of the external noise (from white noise to 1/f^4) and the independent internal noise. Results: The IS-COR achieved higher localization accuracy than the IS for white noise (44.6%±2.2% vs. 22.4%±1.9%). The accuracy difference across models diminished with increasing independent internal noise. The IS-COR executed larger saccades than the IS model (mean amplitude of 5.088±0.035 degrees vs. 2.383±0.033 degrees). The mean saccade amplitudes of the IS-COR also varied with the statistical properties of the external noise, making larger saccades with low-pass noise (mean amplitude of 6.116±0.047 degrees). Conclusion: A novel ideal searcher (IS-COR) that predicts the optimal fixations with the more realistic conditions of inter-saccade correlations can be an important tool for calculating optimal eye movements for image-computable foveated search models and for understanding human eye movements.