In the age of technology, machine learning is an exciting and very promising field of research that will advance science in numerous ways – from improving weather forecasts to predicting cancer progression. The EU-funded DEEP TRANSFER (Deep transfer: generalizing across domains) project worked on developing new algorithms that learn models from data to enhance machine learning. Generally, a central challenge in machine learning is obtaining a sufficient amount of training data to be able to learn an accurate model. The project aimed to overcome this challenge by enabling an algorithm to consider data from a source task as well as data from the target problem when learning a model. It focused on the concept of deep transfer, i.e. the ability to transfer knowledge across very different domains. To achieve its aims the project team developed a novel framework that exploits cutting-edge deep transfer. It relied on a generative model of the world that performs transfer by encouraging the reuse of important, automatically discovered high-level regularities (e.g. transitivity, symmetry, and homophily) across domains. This was tested on three real-word data sets, namely a protein-protein interaction dataset related to yeast, a web domain about computer science departments, and a collection of Twitter data. Compared to other transfer learning approaches and learning from scratch, the team’s approach resulted in significant improvement in both accuracy and run time. It also furthered knowledge in other areas of machine learning, including empirical evaluation of learned models, probabilistic model learning, and statistical relational learning. Beyond the cutting-edge technical advances that it has achieved in the field, the project also published the relevant code and data for several systems, facilitating further research in the field. One particular application where a project’s algorithms proved successful involved analysis of disease interactions. As machine learning becomes a more ubiquitous technology, it is bound to aid in more science-led discoveries and applications.
Deep transfer, machine learning, probabilistic model learning, statistical relational learning, disease interactions