Machine learning's goal is to devise algorithms that improve with experience. Currently, experience is largely defined to be the amount of available data. Unfortunately, acquiring data can be time consuming (e.g., annotating documents), monetarily expensive (e.g., genetic testing), physically invasive (e.g., collecting a tissue sample) or unavailable in sufficient quantities (e.g., data about rarediseases). For some tasks, this makes it challenging to obtain the quantities of data necessary to build a sufficiently accurate predictive model. Machine learning algorithms are applicable to manydomains, but cannot generalize across different domains because of the underlying assumption that the training (used to learn the model) and test (used to evaluate the model) data come from the samedistribution. However, in the real world this is often not the case. People are much more adept at handling this than machines and are even able to reapply knowledge learned in one domain to anentirely different one. Yet standard machine learning approaches are unable to do this. Computationally, the missing link is the ability to discover structural regularities that apply to many different domains, irrespective of their superficial descriptions. This is arguably the biggest gap between current machine learning systems and humans. To address this problem, algorithms must be able to perform deep transfer, which involves generalizing across entirely different domains (i.e., between domains with different objects, classes, properties and relations). Few learning algorithms are able to do this. In this project, we will attempt to develop a well-founded, fully automatic approach to deep transfer that discerns complex structural regularities and determines which of theseproperties are likely to apply to a given target task. Deep transfer offers a fundamentally different and novel paradigm for acquiring experience: exploiting data from other, possibly very different, tasks.