Advanced Deep Learning with Python

Meta-training and meta-testing

The scenarios we described in the Zero-shot learning and One-shot learning sections are referred to as meta-testing phases. In this phase, we leverage the knowledge of a pretrained network and apply it to predict previously unseen labels with the help of only a small support set (or no support set at all). We also have a meta-training phase, where we train a network from scratch in a few-shot context. The authors of Matching Networks for One Shot Learning introduce a meta-training algorithm that closely matches the meta-testing. This is necessary so that we can train the model under the same conditions that we expect it to work in the testing phase. Since we train the network from scratch, the training set (denoted ...