Fig. 12: Architecture backbone, building blocks and elements.
From: Modeling attention and binding in the brain through bidirectional recurrent gating
a Detailed architecture used for the MNIST experiment. For the MNIST model, we use RNN layers in the bottleneck. b Detailed architecture used for the COCO experiment. For the COCO model, we use stride of 2 in convolutional layers for downsampling instead of max-pooling.
