Augmenting machine learning algorithms with the addition of a physics based intelligence prior

Improving the predictive power of machine learning models is one of the
greatest challenges to the science of learning. Here we demonstrate with
the simplest of neural networks that the addition of an intelligence prior
can drastically improve the learning capabilities. We outline this simple
mechanism to decrease the number of exposures, and enhance the predictive
power with a number of examples relevant to the study of quantum phase
transitions. We find that guided networks are uniquely capable to identify
key features of quantum phases where unguided models fail, and that while
the mean square error of topologically equivalent models may be commensurate,
the structure of the predictions produced by the models is qualitatively
very different. In many situations where knowledge of a physical system is
available, but direct sampling of the entire phase space is computationally
intractable, this approach offers a superior learning alternative.