Visual Intelligence aims to develop new deep learning methodology which exploit context, dependencies and prior knowledge in deep learning.
Most of current deep learning systems for image analysis depend on individual pixel information, capturing solely neighborhood dependencies via convolutions. This means that the ability to incorporate context and prior knowledge, e.g. about the topology or geometry of objects, 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 spatial and temporal dependencies and processes. In most applications, additional expert knowledge is available but not expressed in a form understandable by a machine. Moreover, in some applications, additional data could be incorporated in the learning process. These data may come from other sources and be related indirectly to the learning task.
Visual Intelligence has proposed new innovative methods for including expert knowledge and additional data in deep learning models. Such results include the following:
• A method for locating key points in mammograms using graph convolutional networks.
• A method which includes contextual information about depth and time-of-day into a segmentation network in marine acoustics.
• A method for combining deep learning models and Hough transform for building detection.
These methods successfully incorporate dependencies and prior knowledge into the deep learning systems. By exploiting these geometric or contextual constraints, the search space of functions which a network aims to estimate can be significantly reduced. This in turn allows for more efficient training, in particular in a setting with only limited available training data, and results in solutions aligned with the task and application.
