Figure 3.

Deep Learning model. (A) An overview of the DL model, which was composed of two parallel convolutional neural networks (CNNs) encoding two scales of visual input to extract high-level representations of an image and predict the corresponding difference map between fixations of 18-month-olds and 30-month-olds. At each fixation point, a 1024-dimensional feature vector was extracted from the convolutional layers. The feature vectors were integrated across trials to represent each toddler’s eye-tracking patterns. These representations were classified with a linear SVM to distinguish the two age groups. (B) The ROC curve of the DL classification. Positive values indicate 30-month-olds; negative values indicate 18-month-olds.