Fig. 4

Comparison between (deep) reinforcement learning/Deep Q learning and supervised deep learning. The mean accuracy of the testing set of 30 images is plotted as a function of training time, which is measured for reinforcement learning in episodes, and for supervised deep learning in epochs. Both methods were trained on the same 30 training set images. Blue circles represent the results of reinforcement/Deep Q learning, with best fit line in red. Accuracy for supervised deep learning is shown as yellow diamonds, with best fit line in green. Whereas supervised deep learning quickly overfits the small training set and thus cannot learn from it in any generalized manner, the Deep Q network is able to learn over time in a way that can generalize to the separate testing set