中科院固体物理研究所知识库

In order to apply a convolutional neural network (CNN) to unseen datasets, a common way is to train a CNN using a pre-trained model on a big dataset by fine-tuning it instead of starting from scratch. How to control the fine-tuning progress to get the desired properties is still a challenging problem. Our key observation is that the visual features of the pre-trained model have rich information and can be explored during the training process. A natural thought is to employ these features and design a control strategy to improve the performance of the transfer learning process. In this paper, a procedural learning framework using the learned low-rank component of the visual features both in the pre-trained model and the training process is proposed to improve the accuracy and generalizability of the CNN. In this framework, we presented an approach to yield independent visualization features (IVFs). We found via robust independent component analysis that the low-rank components of IVFs provided robust features for our framework. Then, we design a Wasserstein regularization to control the transportation of the distribution of IVFs from a pre-trained model to a final model via the Wasserstein distance. The experiments on the Cifar-10 and Cifar-100 datasets via a VGG-style CNN model showed that our method effectively improves the classification results and convergence speed. The basic idea is that exploring visual features can also potentially inspire other topics, such as image detection and reinforcement learning.