The strength of machine learning methods is the ability to learn from data rather than using predefined models. For complex data there is however a need to integrate the best of these two worlds to enable integration of physical or geometrical models, dependencies, and prior knowledge, as well as the exploitation of multiple complex image modalities simultaneously.
Current deep learning systems for image analysis depend on individual pixel information, capturing dependencies solely via the convolution neighborhood.
This means that the ability to incorporate context and prior knowledge, e.g. about topology or boundaries, is limited. The ability to conform to physical models, and principles governing the image data generation and its properties is also limited, including modelling of temporal dependencies and processes. In order to make deep learning based computer vision systems ubiquitous and applicable also for complex, sparsely labelled image data, there is a need for visual intelligence that can easily be adapted to new, non-standard data sources with few labelled training samples.
To develop new methodology to exploiting context, dependencies and prior knowledge in deep learning.
