2B–C). Kernel size only affected the overall performance of the model by around 1%. With a kernel size larger than four, the model stably produces accuracies over 87% (Supplemental Fig. 2B–C). Similarly, filter number only slightly changed the accuracy (Supplemental Fig. 2D). Pooling layer, on the other hand, was crucial after the convolution layer. Not only did it downsize the information to reduce computation time, but it also increased the classification accuracy.70 Network without pooling usually had a substantial decrease of performance which was usually caused by the propagation of local features to the neighbors.71 So, including the pooling operation can shrink the feature map while still preserving key information required for classification.38, 72, 73 We investigated two popular pooling methods, average pooling and max pooling, each with its own strengths and weaknesses.38, 74 The evaluation of the model accuracy with different pooling methods showed that max pooling outperformed average pooling despite the percentage of human reads (Supplemental Fig. 2E). Finally, we showed that instead of classical stochastic gradient descent procedure, the Adam optimization algorithm,75–77 performed the best (Supplemental Fig. 2F).
