We aim to tackle these challenges via a representation that
jointly predicts winds, temperature, pressure, and dew point
across space and time. The proposed architecture combines
a bottom-up predictor for each individual variable with a
top-down deep belief network that models the joint statistical
relationships. Another key component in the framework
is a data-driven kernel, based on a similarity function that
is learned automatically from the data. The kernel is used
to impose long-range dependencies across space and to ensure
that the inferences respect natural laws.We present an
ecient procedure for combining inferences from separate
predictors of local phenomena while considering constraints
imposed by the deep belief network such that the predictions
respect the natural regularities expected with the large-scale
phenomena.