We first show, as proof of concept in a toy example, that the simple decoder architecture in Figure 2 is able

/Decoder accurately predicts 2D and 3D ellipses from genotype data with 1000 training samples and 32,768 output dimensions