Fig. 4: Contribution of top-down computations to illusory contour responses.
From: Top-down perceptual inference shaping the activity of early visual cortex
a Illustration of the illusory contour experiment. Left: receptive field of an example model Z1 neuron. Middle: receptive field-aligned Kanizsa square stimulus. Red border indicates the actual boundary of the stimulus. Arrow: direction of shifts of the stimulus relative to the neural receptive field. Right: real square (`Line'), Kanizsa square (`Illusory') and incongruent (`Rotated') stimuli. b Mean responses of the Z1 neuron in (a) to 500 presentations of the three stimuli as a function of stimulus shift. Shaded regions (invisibly small): s.e.m. c Same as (b) for a selected unit in the V1 of a macaque monkey. Number of trials is unknown. Reproduced from48 with permission (PNAS, Copyright (2001) National Academy of Sciences, U.S.A.). d Ratio of mean Z1 responses to `Illusory' and `Line' stimuli and to `Rotated' and `Line' stimuli, respectively, at the `Line' response peak, for the largest stimulus sizes that fit into the patch, for the analyzed Z1 population (n = 114 = (57 central, localized, medium wavelength Z1 filters) × (upright and upside-down stimulus orientations), Methods; center line, median; box limits, upper and lower quartiles; whiskers, 1.5 × interquartile range). Mean peak ratios: 0.70 (`Illusory'), 0.23 (`Rotated'), the latter being significantly less than the former (one-sided paired two-sample t-test, t(df = 113) = 4.3, p = 1.6 × 10−5, \({d}_{e}^{{\prime} }=0.39\), 95% confidence interval = [0.29, ∞]). *p < 0.05, **p < 0.01, ***p < 0.001; n.s., p≥0.05 in this and all subsequent Figures. e Same as (b) for the `Line' and `Illusory' stimuli, together with linear responses of the Z1 neuron (dashed lines). Magnitude of the linear response to the `Line' stimulus was scaled to the mean response peak to the `Line' stimulus. Shaded regions (invisibly small): s.e.m. f Same as (e) for the `Line' and `Rotated' stimuli. g Differences between the mean and linear responses to the `Illusory' stimulus (center line, median; box limits, upper and lower quartiles; whiskers, 1.5 × interquartile range). Left: Same restricted Z1 population of TDVAE in three different conditions (n = 9 each; see text, Methods): intact inference in Z1 (mean model-linear response difference (0.076) is positive: one-sided one-sample t-test against 0: t(df = 8) = 4.7, p = 0.00076, 95% confidence interval = [0.046, ∞]), inference without stimulus-specific information at Z2 (mean model-linear response difference (−0.068) is not significant: two-sided one-sample t-test against 0: t(df = 8) = − 0.71, p = 0.50, 95% confidence interval = [−0.29, 0.15]), inference with Z2 clamped to zero (mean model-linear response difference (0.014) is significant: one-sided one-sample t-test against 0: t(df = 8) = 2.9, p = 0.011, 95% confidence interval = [0.0049, ∞], but smaller than TDVAE with intact inference: one-sided paired two-sample t-test: t(df = 8) = − 4.2, p = 0.0015, \({d}_{e}^{{\prime} }=2.07\), 95% confidence interval = [−∞, −0.034]). Right: Illusory responses in two control models: shallow-VAE (n = 29; mean model-linear response difference (0.011) is positive: one-sided one-sample t-test against 0: t(df = 28) = 3.2, p = 0.0016, 95% confidence interval = [0.0050, ∞], but smaller than TDVAE with intact inference: one-sided independent two-sample t-test: t(df = 36) = − 6.1, p = 2.2 × 10−7, \({d}_{e}^{{\prime} }=2.06\), 95% confidence interval = [−∞, −0.047]), and feed-forward goal-directed model (n = 40; mean model-linear response difference (−0.00030) is not significant: two-sided one-sample t-test against 0: t(df = 39) = −0.37, p = 0.71, 95% confidence interval = [−0.0019, 0.0013]). Filled circle: significant boosting or suppression. Hollow circle: no significant effect. Square: deterministic model. Source data are provided as a Source Data file.
