Fig. 1

Meta-generalization experiment on 1D sinusoids. CNPs (left) and MAML (right) meta-trained on sinusoids with frequency \(\rho =1\)  generalize to similar test functions. However, changing the frequency to \(\rho = 1.5\)  leads to catastrophic loss of generalization. Quantitative results are shown in Table F.1. This figure is inspired by Finn et al. [3]