Once we see that the features in our training data allow for okay training, we implement a simple type of stochastic model parameterized as a multivariate Gaussian, where each dimension belongs to a feature in the output space (the space of object-catalog features).
Call the input (truth-catalog) features x and the output (object-catalog) features y. The central challenge of this task is defining a loss function that captures the distance between the current network distribution q(y|x; theta) where theta refers to the Gaussian parameters and the target distribution p(y|x).
Once we see that the features in our training data allow for okay training, we implement a simple type of stochastic model parameterized as a multivariate Gaussian, where each dimension belongs to a feature in the output space (the space of object-catalog features).
Call the input (truth-catalog) features
xand the output (object-catalog) featuresy. The central challenge of this task is defining a loss function that captures the distance between the current network distributionq(y|x; theta)wherethetarefers to the Gaussian parameters and the target distributionp(y|x).