Fig. 5: Model displays perceptual learning effect and predicts human learning outcomes.

a DNN model schematic. The model compares bottom-up features with the “state” module representing prior knowledge, produces top-down conditioning features, and then produces a final output, which is then used to update the model state. For details, see “Methods” and Supplementary Fig. 10. b Model accuracy on 1000 test sequences (sequence length: 630, including 210 unique Mooney images) constructed from grayscale ImageNet 1k images and their Mooney image counterparts. The repetition effect was evaluated on the same Mooney images presented twice in a sequence without matching grayscale images (sequence length: 420). The plot shows the distribution of aggregate phase performance across different synthetic image presentation orders (n = 1000 image presentation orders). **** indicates p < 0.00001, Mann–Whitney test. The central white dot of each violin plot represents the median, the gray vertical bar represents the interquartile range (25th to 75th percentiles), the violin plot bounds represent the minima and maxima, and the plot curvature represents the density estimate of the data distribution. Detailed model performance for “pre”, “post”, and “gray” conditions is plotted in Supplementary Fig. 11c, broken down by the position of a trial within the long image sequence. c Model learning performance as compared to human subjects, where the model was presented with identical image sequences as human subjects. Whiskers show min and max of accuracy, box sides show 25 and 75 percentile, center line shows median accuracy (n = 219 images). d Image recognition error pattern alignment between human subjects (H- > H), between model fed with difference sequences of the same images (M- > M), and between model and human subjects with matching image presentation (M- > H matching) and non-matching image presentation order (M-H non-matching), measured by AUROC. Bar height shows median value across measurements and error bars indicate 95% CI of AUROC. Dashed line indicates chance level AUROC. *** indicates statistical significance above chance (p < 0.0005, one-sided t-test). e Human learning outcome prediction. On the x-axis, numbers refer to model layer and CLS, Logits refer to model’s representation following its last layer. Black horizontal bar at the top indicates significant difference in prediction when using model features from pre- vs. post-phase (p < 0.05, two-sided t test with FWE correction, n = 12 subjects). Blue, orange, and gray horizontal bars show significant prediction as compared to the chance level (p < 0.05, two-sided t test with FWE correction, n = 12 subjects). Individual image learning outcomes are significantly predicted using the model’s grayscale image representation starting from the second layer onward. Center line shows mean AUROC across subjects, with shaded areas show 95% CI across subjects (n = 12 subjects).