Skip to main content
. 2020 Jun 9;7:63. doi: 10.3389/frobt.2020.00063

Figure 6.

Figure 6

(A) F1 score for classification on the CIFAR-10 dataset with DCH with and without the HIL, as a function of the number of iterations of training of the DCH network. (B) F1 score for classification on the CIFAR-10 dataset with DCH with and without the HIL, as a function of the Hamming Distance for classification. The networks are fully trained to the end point shown in subplot (A). (C) F1 score for classification on the CIFAR-10 dataset with DTQ with and without the HIL, as a function of the number of iterations of training of the DTQ network. (D) F1 score for classification on the CIFAR-10 dataset with DTQ with and without the HIL, as a function of the Hamming Distance for classification. The networks are fully trained to the end point shown in subplot (C). (E) F1 score for classification on the CIFAR-10 dataset with DQN with and without the HIL, as a function of the number of iterations of training of the DQN network. (F) F1 score for classification on the CIFAR-10 dataset with DQN with and without the HIL, as a function of the Hamming Distance for classification. The networks are fully trained to the end point shown in subplot (E). Baseline networks are shown in blue, while the same network with a HIL appended to the end is shown in yellow. Note that in the left column of subplots, the Hamming Distance for classification is set to 2 for inlier/outlier count. The left column of results show that the HIL boosts the speed at which the network trains, achieving a higher performance in far fewer iterations of expensive network training. As the HIL adds negligible overhead in memory/computation time, there is no downside to using a HIL. The right column of results show that the HIL prevents the need for searching for an optimal Hamming Distance threshold to classify with, as it supercedes peak performance of the network right away for the lowest possible distance thresholds. After peak performance of the baseline network, larger Hamming Distance thresholds eventually decay to the performance of the baseline.