Backpropagation Network‐Based Contrastive Learning for Unsupervised Domain Adaptation

ABSTRACT This study introduces a new method for domain adaptation for image classification tasks that aims to improve the model's performance on a target domain after being trained on a source domain. The method utilizes an unsupervised contrastive learning algorithm for Backpropagation (BP) neural networks and focuses on aligning category representations based on the semantic distribution of samples. The proposed method stands apart from traditional adversarial training by not accounting for variability in semantic representation in image examples. The proposed method supplements a contrastive loss function at the sample level to improve feature alignment between source and target domains. The experiments confirm the effectiveness of the proposed method on a variety of datasets in both classification perspective and domain adaptation perspective, and the proposed method shows substantial classification accuracy and domain adaptation performance from previous methods. Additionally, an upper bound of contrasting loss was derived to make the computational space efficient to conduct and manage many similar sample pairs and dissimilar sample pairs efficiently. Future works could utilize extensions of the current framework in many ways, thus making it future work to conduct research on real‐time domain adaptation, robustness in environments with noise, or multi‐modal data for better adaptation for many sources of reference to account for more adaptation across multiple domains.