基于多轮修正噪声标签的神经网络分类框架-《计算机技术与发展》

文章信息/Info

Title:: A Neural Network Classification Framework Based on Calibrating Noisy Labels in Multi-round

Author(s):: WANG Xue-gang; WANG Yu-feng; School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China

Keywords:: noise labels; label transition matrix; weighted average noise rate; multi-round calibration; neural network classification framework.

摘要:: 利用大规模的带标签数据集训练神经网络在分类任务中表现出色,但是实际使用的数据集中通常包含噪声标签从而使得分类网络的性能变差。为了克服噪声标签的不利影响,提出了一种基于多轮修正噪声标签的神经网络分类框架。该方法在每一轮修正中均更新训练的网络参数并修正当前训练数据中的噪声标签,修正后的数据集用于下一轮训练和修正。具体而言,在每一轮修正中首先利用本轮的数据集训练网络,并利用“ 锚点样本” 的网络预测值估计数据集的标签转移矩阵;然后计算数据集的加权平均噪声率;之后结合加权平均噪声率和数据样本的训练损失值依据“ 小损失冶原则筛选出噪声标签;最后利用标签转移矩阵和网络预测值对噪声标签进行自适应修正。经多轮修正可有效地降低数据集的噪声水平,从而使得训练出的分类网络更加准确。在多个真实数据集上的实验结果表明,该方法与现有的方案相比有较大的性能提升。

Abstract:: Neural networks trained with large-scale labeled datasets have shown excellent performance in classification tasks. However,the datasets used for practical tasks often contain noisy labels,which will make the performance of the classification network worse. Inorder to overcome the adverse effects of noisy labels,　we propose a neural network classification framework based on calibrating the noisylabels in multi-round,which updates the trained network parameters and calibrates the noisy labels in the current training data in eachround of calibration,the calibrated dataset is used for the next round of training and calibration. Specifically,in each round,the predictiveresults of some selective anchor samples is utilized to estimate the label transition matrix of the current data,which is then used to infer theweighted average noise rate. Then,through exploiting the " small loss" principle,the noisy data samples are chosen by considering theweighted average noise rate and the trained loss value of data sample. Finally,the noisy labels are calibrated through combining the labeltransition matrix and predictive result of the trained network adaptively. After multiple rounds of calibration,the noisy level of the datasetis significantly reduced,based on which more accurate classification result can be trained. Experimental results on multiple real datasetsshow that the proposed method has a large performance improvement compared with existing schemes.