Fig. 3: Signatures of learning a texture representation.
From: Top-down perceptual inference shaping the activity of early visual cortex
a Mean responses of one example neuron from Z2 and one example neuron from Z1 (rows two and three, respectively) to three example images from three texture families each (top). Bottom: Variability of mean responses of the same two neurons across 4000 images in each of the 15 texture families (colors; center lines: means, whiskers: extrema). b Same as (a) top in V1 and V2 recordings from macaque monkeys. c Left: Partitioning single-unit response variances into variance across images from different families, variance across images within families and a residual component (across stimulus repetitions) with nested ANOVA both in Z2 and Z1 (n = 100 neurons, dots). Insets: Distribution of the sum of the first two components of variance. Right: Distribution of the ratio of the first two components. Geometric mean of variance ratios (triangles) is significantly smaller in Z1 (0.064) than in Z2 (2.097) (one-sided independent two-sample t-test in the log domain, t(df = 198) = − 32.7, p = 6.5 × 10−82, \({d}_{e}^{{\prime} }=4.7\), 95% confidence interval = [−∞, −3.3]). Here and later in the paper, \({d}_{e}^{{\prime} }\) denotes the average sd discriminability index. d Same as (c) in 102 V1 and 103 V2 units from a pool of 13 macaque monkeys. e Decoding texture family (red) and within-family stimulus identity (black) from the same 100 Z1 and Z2 units (horizontal axis and vertical axis, respectively) as in (c). Dot sizes: decoding on randomly selected subpopulations (n = 1, 3, 10, 30, 100). Solid line: Chance performance. Dot centers: means; error bars: 95% confidence intervals of bootstrapping over included neurons and partitionings. f Same as (e) from single unit V1 and V2 macaque recordings. g Phase scrambling-induced modulation of the magnitude of mean responses to texture images in the same 100 Z2 (top) and Z1 (bottom) neurons as in (c). Mean modulation index (triangles) is significantly smaller in Z1 (−0.062) than in Z2 (0.027) (one-sided independent two-sample t-test, t(df = 198) = −12.4, p = 8.6 × 10−27, \({d}_{e}^{{\prime} }=1.77\), 95% confidence interval = [−∞, −0.077]). h Same as (g) from V1 and V2 recordings in macaques. Experimental data panels are reproduced with permission from PNAS (b−f 17) and SNCSC (h16). Source data are provided as a Source Data file.
