Kernels with peripheral vison: weights of different significance
Zixuan He · Journal of Physics Conference Series · 2023
Abstract The visual field of humankind which is generally composed of central and peripheral vison has different acuity at different spots, while in a convolution layer in convolution neural networks, different weights in a kernel are of the same significance. In order to attach weights with decreasing significance from the center of the kernel to its edge, this paper proposes kernels with peripheral vison, imitating the variable acuity of the human’s visual field. It is achieved by adding a module to the loss function, and a comparation between its implements in Residual Network (ResNet) and GoogLeNet on CIFAR100 is provided in this paper. With this method, kernels whose weights have different sensitivity of the input data are created. It will consume insignificant energy, since it is only functioned when training. The improvements in valid accuracy of ResNet and GoogLeNet are 0.6 percent and 2 percent. It can be easily attached to most structures, improving the accuracy of these structures.