Figure 3
Adaptation of the DirectRanker architecture to use convolutional layers to extract additional features from the inputs. The two convolutional networks share their parameters and architecture. After the convolutional layers, a fully connected layer is used the further reduce the extracted features before they are concatenated with the original ones. The concatenated features are then fed into the original DirectRanker model12. During training, a batch normalization layer37 is used before the extracted features are added to the primary ones to prevent overfitting.
