Efficient Learning of Domain-invariant Image Representations

摘要: Abstract: We present an algorithm that learns representations which explicitly compensate for domain mismatch and can be efficiently realized as linear classifiers. Specifically, we form a transformation maps features from the target (test) to source (training) part of training classifier. optimize both classifier parameters jointly, introduce efficient cost function based on misclassification loss. Our method combines several previously unavailable in single algorithm: multi-class adaptation through representation learning, ability map across heterogeneous feature spaces, scalability large datasets. experiments image datasets demonstrate improved accuracy computational advantages compared previous approaches.