A Model Zoo Details Zoo Input Channels Parameters Population Size model zoos share one general CNN architecture, MNIST 1 2464 1000 SVHN 1 2464 1000 choices for each of the population

The model zoos are generated following Table 5: Model zoo overview. Figure 8: Schematic of the auto-encoder architecture to learn hyper-representations. Hyper-representations are learned with an autoencoder based on multi-head self-attention. The architecture is outlined in Figure 8. Convolutional and fully connected neurons are embedded to token embeddings of dimension d A learned compression token (CLS) is appended to the sequence of token embeddings. The sequence is passed through another stack of multi-head self-attention, which is symmetric to the encoder.