Fig. 1: Minimum convergence sample estimation can be used to approximate the number of labels required for generalizable performance.
From: Autoencoders for sample size estimation for fully connected neural network classifiers
a A fully connected network is trained on labeled data, and tested on a unlabeled data. Generalizability Performance is measured via AUC. Minimum convergence sample (MCS) reflects the minimum number of labeled samples required for a fully connected network to start generalizing. b An autoencoder with a similar structure as the fully connected network is trained on unlabeled data and the loss function measures how generalizable the FCN is. Minimum convergence sample estimate (MCSE) approximates the minimum convergence sample (MCS).
