Figure 1.

Convolutional Neural Network Training

Images were taken from pluripotent and differentiating 46C mESCs and fed into a ResNet50. On the left panels it can be seen how colonies acquire a spindle shape as time progresses. Scale bars, 100 μm. Early changes are present at 2 h after the onset of differentiation. In the center column, each panel presents the training and validation accuracy, as well as the training and validation loss. As time from the onset of differentiation shortens, the training becomes less confident. Although an acceptable training is reached at 2 h, validation accuracy does not reach a value close to 1. No training is obtained at time zero. On the right columns, confusion matrix graph shows the results of testing the CNN on 100 independent images. Accuracy is at the top at 6, 12, and 24 h, falling to 0.97 at 2 h.