Fig. 2: Experimental results on image classification tasks. | Communications Engineering

Fig. 2: Experimental results on image classification tasks.

From: Improving the robustness of analog deep neural networks through a Bayes-optimized noise injection approach

Fig. 2

a Experiment results on the MNIST dataset. b Experiment results on the CIFAR-10 dataset. The left part in each panels a, b is a schematic demonstration of the task itself. The results are presented by curve charts and bar charts. The curve chart compares the prediction accuracy of our methods (BayesFT-Ga, BayesFT-La, and BayesFT-Do) and the baseline methods (ERM, FTNA, AWP, and ReRam-V) at different resistance variation (σ = 0–1.5). The shaded areas are confidence intervals. The bar charts with confidence intervals show the results of statistical tests (i.e., run the task 20 times and compare the accuracy) at a specific σ setting. The horizontal line above the bars indicates the statistical difference in the performance of our methods compared to the baseline methods (if the difference is significant, it will be marked by the *** symbol. Otherwise, the p value will be displayed). Specific σ values are selected as they can present challenging environments.

Back to article page