Skip to main content
. 2019 Dec 2;9:18113. doi: 10.1038/s41598-019-54653-6

Figure 1.

Figure 1

Illustrating the VAE for visual fields. (A) The top frame demonstrates the structure of the VAE, which is a dual-mapping from the original visual field to the latent features and then back to a reconstructed visual field. (B) The bottom frame provides details on the network architecture. There are four types of layers: convolutional (C), resizing (R), fully connected (F), and de-convolutional (D). The sizes at the bottom of each image reflect the transformed dimensions after each layer. For each layer, the kernel is 3 × 3 with a stride of 2. The activation function for all layers is the rectified non-linear unit, except for F1 and D3, which have the identity and sigmoid activations, respectively. The latent dimension, L, is a user specified parameter.