Fig. 2
From: Brain-guided convolutional neural networks reveal task-specific representations in scene processing
Illustration of our brain-guided CNN and neural similarity pooling operation. Top: the network consisted of four convolutional layers, each being guided by neural responses sampled from one of four different sequential time points from ERP data measured across all images and both tasks. The output of the network consisted of a single decision node that predicted which task an observer was performing on the image. Bottom: we replaced the typical max-pooling operation with a neural similarity pooling operation that selected filter outputs that most closely matched the neural response elicited by each image (see "Methods" section for further detail).
