MNIST数据库
计算机科学
人工智能
机器学习
滤波器(信号处理)
图像(数学)
模式识别(心理学)
字错误率
深度学习
计算机视觉
作者
Hao Xu,Hui Xiao,Huazheng Hao,Li Dong,Xiaojie Qiu,Chengbin Peng
标识
DOI:10.1016/j.knosys.2022.110166
摘要
Semi-supervised learning frameworks usually adopt mutual learning approaches with multiple submodels to learn from different perspectives. Usually, a high threshold is used to filter out numerous low-confidence predictions for unlabeled data to avoid transferring erroneous pseudo-labels between the submodels. However, such filtering cannot fully exploit unlabeled data with low prediction confidence. In this study, we propose a mutual learning framework based on pseudo-negative labels to overcome this problem. Negative labels are those to which a corresponding data item does not belong. In each iteration, one submodel generates pseudo-negative labels for each data item, and the other submodel learns from these labels. The roles of the two submodels interchange after each iteration until convergence. The dual model can improve its prediction ability by reducing the prediction probability of the pseudo-negative labels. We also propose a mechanism for selecting a few pseudo-negative labels to be fed into the submodels. In experiments, our framework achieved state-of-the-art results on several main benchmarks. Specifically, with our framework, the error rates of the 13-layer CNN model were 9.35% and 7.94% for CIFAR-10 with 1000 and 4000 labels, respectively. In addition, our framework has an error rate of 0.81% for the non-augmented MNIST with only 20 labels, which is much smaller than that of the other approaches. Our approach also demonstrated a significant performance improvement in domain adaptation.
科研通智能强力驱动
Strongly Powered by AbleSci AI