Figure 2.

Architectures of CNPs and ANPs. (a) Components of a CNP model: the encoder of CNPs is composed of a deterministic path to generate a representation r; the decoder of CNPs uses the r and target $x_T$ to produce the mean and the Std. (b) Components of an ANP model: the encoder of ANPs is composed of a deterministic path to generate a representation $r^{*}$ and a latent path to generate latent variable z; the decoder of ANPs uses the $r^{*}$, z, and target $x_T$ to produce the mean and the Std.