Meta-Transfer Learning for Few-Shot Learning

In Figure S1, we present the 4CONV architecture for feature extractor Θ, as illustrated in Section 5.1 “Network architecture” of the main paper. In Figure S2, we present the other architecture – ResNet12. Figure S2(a) shows the details of a single residual block and Figure S2(b) shows the whole network consisting of four residual blocks and a mean-pooling layer. The input of Θ is the 3-channel RGB image, and the output is the 512-dimensional feature vector. a = 0.1 is set for all leakyReLU activation functions in ResNet-12.