基于可调节判别器的领域适应

当前基于对抗学习的领域适应(DA)对目标样本的适应性较差,导致目标域的预测精确度较低,为此提出基于可调节判别器的领域适应(A-DADA)算法.首先,利用两个判别器分类概率的距离作为权重应用到目标域对抗训练损失函数中,旨在减少已对齐目标样本对抗训练的次数同时增加未对齐目标样本的对抗训练次数;其次,将平方熵损失函数作为最小...

Domain adaptation based on adjustable discriminator

The current domain adaptation (DA) based on adversarial learning has poor adaptability to target samples, lower prediction accuracy of the target domain is caused. For this reason, this paper proposes a domain adaptation (A-DADA) algorithm based on an adjustable discriminator. First, the distance between the classification probabilities of the two discriminators is used as the weight to apply to the target domain confrontation training loss function, aiming to reduce the number of confrontation training times for aligned target samples and increase the number of confrontation training times for unaligned target samples. Secondly, square entropy loss function is used as the minimum entropy loss function to improve the class balance. Finally, the Office-31 datasetis used to compare the algorithm with JAN and other algorithms, the experimental results show that compared with the JAN algorithm, the average accuracy of the algorithm is increased by 2.5%.