This step focuses on the minimization of the fully-connected layer size. Starting from 512 neurons in the CNN B, we have reduced the number of neurons by a factor of 2 until only 2 neurons are left in the CNN L (see Table 3). We also evaluate the performance of the CNN M where there is no fully-connected layer at all (in this case, the outputs of the last convolutional layer are directly connected with 2 neurons of the Softmax layer). The obtained results are presented in Fig. 4.
This time, the difference between CNNs is less significant than in the 2 first optimization steps. Apparently, the size of the fullyconnected layer is less influential on the final accuracy than the number and the width of the convolutional layers.
However, we can observe that CNNs B, E–I reach the selection threshold of 97.5% classification accuracy with respect to standard deviations, while the performances of the CNNs J–M are below the selection threshold on the validation set. Hence, the CNN I is selected after the last optimization step.
This step focuses on the minimization of the fully-connected layer size. Starting from 512 neurons in the CNN B, we have reduced the number of neurons by a factor of 2 until only 2 neurons are left in the CNN L (see Table 3). We also evaluate the performance of the CNN M where there is no fully-connected layer at all (in this case, the outputs of the last convolutional layer are directly connected with 2 neurons of the Softmax layer). The obtained results are presented in Fig. 4. This time, the difference between CNNs is less significant than in the 2 first optimization steps. Apparently, the size of the fullyconnected layer is less influential on the final accuracy than the number and the width of the convolutional layers. However, we can observe that CNNs B, E–I reach the selection threshold of 97.5% classification accuracy with respect to standard deviations, while the performances of the CNNs J–M are below the selection threshold on the validation set. Hence, the CNN I is selected after the last optimization step.
การแปล กรุณารอสักครู่..
