Emerging data resources in the big data era have created new opportunities to break the main bottleneck of machine learning systems; namely, their reliance on human annotated training data. Generalized information adaptation provides mechanisms for exploiting freely available auxiliary data resources to reduce human effort and improve the autonomy of machine learning.

In this talk, I will first briefly introduce the problem of generalized information adaptation, and then discuss a special type of information adaptation, zero-shot learning, which transfers knowledge across classes to recognize new objects. In particular, I will present zero-shot learning methodologies that exploit both text and image resources to build connections across various class categories. Promising empirical results are reported.