Improved generalization performance of convolutional neural networks with LossDA

Abstract In recent years, convolutional neural networks (CNNs) have been used in many fields. Nowadays, CNNs a high learning capability, and this capability is accompanied by more complex model architecture. Complex architectures allow to learn data features, but such process tends reduce the training model’s ability generalize unknown data, may be associated with problems of overfitting. Although regularization methods proposed, as augmentation, batch normalization, Dropout, research on improving generalization performance still common concern robust CNNs. paper, we propose dynamically controllable adjustment method, which call LossDA, that embeds disturbance variable fully-connected layer. The trend kept consistent loss, while magnitude can preset adapt different models. Through dynamic adjustment, adaptively adjusted. whole improve helping suppress To evaluate paper conducts comparative experiments MNIST, FashionMNIST, CIFAR-10, Cats_vs_Dogs, miniImagenet datasets. experimental results show method Light Transfer (InceptionResNet, VGG19, ResNet50, InceptionV3). average maximum improvement accuracy 4.62%, F1 3.99%, Recall 4.69%. 4.17%, 5.64%, 4.05%.