PMC

Convolution of two exponential functions

Functions of Deformable convolution and Deformable Pooling

Network training loss of single-scale convolution kernel. (a) The important inflection points of network loss convergence. The loss functions of different convolution kernels decrease rapidly before 1,000 iterations. Additionally, the loss function curves of all convolution kernels tend to be flat before 3,000 rounds. (b) The loss function performance of different convolution kernels after multi-round iteration. Long-shape and small-size convolution kernels are suitable for feature extraction of ERP tasks.

Components of constructed convolution layers and overall CNN_5layers construction. a The picture on the top shows details of the constructed convolution layer, which contains a convolution operation, a batch normalization and a leaky rectified linear unit in turn. The convolution kernel size is 3 ×1 and the output channel is 120. The simplified diagram is on the bottom. b The picture shows the overall CNN_5layers schema, including five convolution layers, five maximum pooling layers, and two fully-connected layers. All activation functions in CNN_5layers were LeakyReLU. One dropout layer which drops 30 percent was performed after each pooling operation and the first fully-connected layer