Deep Subdomain Adaptation Network for Image Classification

For a target task where the labeled data are unavailable, domain adaptation can transfer learner from different source domain. Previous deep methods mainly learn global shift, i.e., align and distributions without considering relationships between two subdomains within same category of domains, leading to unsatisfying learning performance capturing fine-grained information. Recently, more researchers pay attention subdomain that focuses on accurately aligning relevant subdomains. However, most them adversarial contain several loss functions converge slowly. Based this, we present network (DSAN) learns by domain-specific layer activations across domains based local maximum mean discrepancy (LMMD). Our DSAN is very simple but effective, which does not need training converges fast. The be achieved easily with feedforward models extending LMMD loss, trained efficiently via backpropagation. Experiments demonstrate achieve remarkable results both object recognition tasks digit classification tasks. code will available at https://github.com/easezyc/deep-transfer-learning.