ACDC: Online unsupervised cross-domain adaptation

We consider the problem of online unsupervised cross-domain adaptation, where two independent but related data streams with different feature spaces – a fully labeled source stream and an unlabeled target are learned together. Unique characteristics challenges such as covariate shift, asynchronous concept drifts, contrasting throughput arise. propose ACDC, adversarial domain adaptation framework that handles multiple complete self-evolving neural network structure reacts to these defiances. ACDC encapsulates three modules into single model: A denoising autoencoder extracts features, module performs conversion, estimator learns predicts stream. is flexible expandable little hyper-parameter tunability. Our experimental results under prequential test-then-train protocol indicate improvement in accuracy over baseline methods, achieving more than 10% increase some cases. • novel transfer. autonomous can grow prune nodes on all training phases. The usage domain-adversarial bias–variance trade-off adapt discriminator. Domain-adversarial learning configuration. Source-code made publicly available for further study.