Machine learning is the science of generating computer programs from data to perform tasks without being explicitly programmed to do so. Machine learning algorithms are used in a wide variety of applications. You probably use a service based on machine learning numerous times a day without knowing it. In the last decade alone, machine learning has contributed to several break-throughs in technology, ranging from speech and language recognition, understanding of biomedical processes and more effective financial services. Many researchers also believe that machine learning is a promising path towards human-level intelligence.
The machine learning group at NEC Laboratories Europe is focusing on the development of machine learning approaches for multi-modal and relational data. Relational or graph-structured data is common in numerous application domains and explicitly represents both the features of data points and their various relationships and interdependencies. For instance, in the biomedical domain, it is common to model genes, drugs, and side-effects as nodes, and their various types of interactions as edges in a graph. In the financial domain, one wants to understand interactions and transactions between banks, companies, and individuals to understand and anticipate the systems’ complex behavior. Generally, we are interested in the problems of the broader field of geometric deep learning that deals with irregular input data for which a direct application of standard (deep) learning methods is challenging.
The research group is also working with multi-modal data, that is, data where each entity is associated with heterogeneous attribute types capturing visual, sequential, and numerical information. For instance, the group is working on top-tier research in natural language processing that allows machines to read, write, and participate in complex interactions through language.
We aim to make our learning algorithms as versatile as possible by developing continual learning approaches allowing them to continuously adapt to different but related problems. This makes the resulting methods more robust to changes in the input data and allows them to better generalize to new learning tasks. We are also working on making machine learning methods more interpretable by combining discrete/symbolic with continuous/neural learning paradigms.