Fig 5. Transfer learning on new bacterial species.

The accuracy of Pseudomonas classification with convolutional neural networks trained in different ways. “Vibrio” indicates training on images of Vibrio bacteria, “Pseudomonas” indicates training on the small Pseudomonas image dataset, and “Transfer” indicates using the Vibrio-derived network weights as the starting point for training on Pseudomonas images. For training only on Vibrio images, the different data points come from random weight initialization, random data ordering, and random augmentation. For training only on Pseudomonas images, and for transfer learning, the different data points are from random train/test splits of the Pseudomonas data.