The developers of Google have extended Tensorflow with the function of so-called Neural Structured Learning, such as the team in his blog telling. This should allow experienced as well as inexperienced developers to be able to train neural networks based on structured input signals.
According to the announcement, the system can be used to create robust models for image recognition, speech processing or general prediction. The input signals can be represented in graphs or other structures. The idea of the learning carried out on it is to evaluate the relations of the thus held data to each other, instead of just using individual input data as before. The concrete procedure should be suitable for any kind of neural networks.
Work on the now available framework is based on research by Google (PDF). According to this, learning on structured data primarily helps to increase the accuracy of models if comparatively few well-characterized output data are available for learning.
As possible output data, the team includes techniques such as the Knowledge Graph, as well as medical records, genetic data, and even multi-modal relations such as text-image pairs. According to their own data, the team at Google is already using this new type of machine learning to improve the performance of certain models. more details for the Neural Structured Learning can be found on the website of Tensorflow, where the team also a tutorial for use the technology provides. The code is found on Github,