Learning to learn

Contact person: prof. dr. Lambert Schomaker

Current methods in machine learning are still unsuitable for many real-life problems. Examples are the recognition of handwriting from undefined scripts and languages in massive collections of historical documents, or dealing with household objects in service robotics at home. All statistical machine learning techniques require large amounts of examples. The manual labeling of the ground truth of image sections or video sequences needs to be replicated for each task, due to the extraordinary variation in patterns. Therefore, it becomes increasingly important to have systems that can learn future tasks ever easier because of their existing skills, as opposed to starting from scratch with a laborious training effort on both the existing tasks plus a newly added task, as is mostly done today. Just as an agent (hero) in the game of Minecraft, an intelligent system should be able to recognize the similarity between a new challenging condition and problems that were solved earlier in its life. Using current computing power, it becomes possible to realize metalearning, where a system is gradually trained to find optimal meta parameters and control parameters for the training of the actual new tasks and skills using deep learning. Today, this is left to human researchers who control the training schedule. This is expensive, the results are highly dependent on human intelligence and often difficult to replicate.

Only if a machine can be trained to find optimal neural-network architectures, select suitable data subsets and determine optimal control-parameter values we can demonstrate that an autonomous mode of artificial intelligence has been reached. A continuously learning autonomous system should be able to master increasingly complex problems, just as the human player of Minecraft or a robot dealing with maintenance problems in a factory.