关键词: 深度学习, 深度神经网络, 图像分类, 标签噪声, 元学习

Abstract: With the continuous progress in the field of computer vision, image classification techniques are becoming more and more mature and are gradually used in various fields. Image classification requires the collection of a large number of images for model training and optimization, but the image collection process inevitably brings noisy labels. To cope with this challenge, robust classification methods have emerged. However, the setting of hyperparameters in the current robust classification methods needs to be adjusted manually, which brings a lot of loss in human and material resources. Therefore, in this paper propose Meta Hyperparameter Adjuster (MHA), which adopts a two-layer nested loop optimization method to adaptively learn noise-aware hyperparameter combinations, and is called Pseudo-label adaptive learning algorithm which based on meta learning (, Meta-FPL). In addition, in order to solve the problem that the backpropagation process in the meta training phase consumes a large amount of GPU arithmetic, the Select the activation metamodel layer strategy (SAML) is proposed, which restricts part of the metamodel layer by comparing the magnitude of the average gradient of the backpropagation and the meta-gradient in the virtual training phase, and effectively improves the model training. updates, which effectively improves the training efficiency of the model. Experiments were conducted on four benchmark datasets and one real dataset respectively, and the experimental results show that the Meta-FPL algorithm has a higher classification accuracy and the training time of the backpropagation process in the meta-training phase is shortened by 79.52%, which suggests that the Meta-FPL algorithm can effectively improve the classification test accuracy in a shorter training time.

Key words: deep learning, deep neural networks, image classification, label noise, meta-learning