Ideal versus computational observer model performance.

Comparison of model performance for simulations in which there were only one stimulus phase and no eye movements, presented at 4.5° eccentricity and typical human optics without defocus. Ideal observer was implemented using the derivation and formula from Geisler [53]. The analytical computation (filled black circles) is the closed form solution given the two (noiseless) templates, assuming the only source of noise is photon noise (Poisson). The ideal observer by simulation (unfilled black circles) applies the ideal decision for 200 trials per stimulus class, where each trial has no uncertainty other than photon noise. The two ideal observer calculations will converge to the same values with sufficiently large simulations. Computational observer performance is replotted from Fig 6B, where a linear SVM classifier learns the mean and variance of cone absorptions for the two stimulus classes for 200 trials per stimulus class (green line). The same computational observer model performs slightly better with more trials (800 per stimulus class, red line).