Fig. 1: The network layout optimized by GenomeNet-Architect consists of three stages: (i) a stage of stacked convolutional layers, (ii) global average pooling (in the CNN-GAP model) or a stack of recurrent layers (in the CNN-RNN model), and (iii) a fully connected stage.
From: Optimized model architectures for deep learning on genomic data
a The CNN-RNN model feeds the output of the last convolutional layer into a block of recurrent layers. The output of the last recurrent layer is then flattened and fed into a fully connected neural network. b The CNN-GAP model groups the convolutional layers into convolutional blocks. While the output of some of these blocks is skipped (controlled by the “skip ratio” hyperparameter), the network performs global average pooling (GAP) on the remaining blocks and concatenates the result. This is then fed into the fully connected neural network.
