Figure 4.

Transfer learning works. (A) Four example images of each of the additional analyzed classes. (B) Applying t-SNE to the network outputs of the additional data (see also Figure 3B) and coloring the points according to the classes demonstrates separation of new compartments based on features trained for classifying other localizations. (C) Classification accuracy on held-out data (y-axis) for different number of training images (x-axis) for DeepYeast outputs (red) or CellProfiler features (blue) used as inputs to a random forest. The error bars denote a 95% C.I. from 20,000 bootstrap samples.