Generic Methods for Optimization-Based Modeling

"Energy” models for continuous domains can be applied to many problems, but often suffer from high computational expense in training, due to the need to repeatedly minimize the energy function to high accuracy. This paper considers a modified setting, where the model is trained in terms of results after optimization is truncated to a fixed number of iterations. We derive “backpropagating” versions of gradient descent, heavy-ball and LBFGS. These are simple to use, as they require as input only routines to compute the gradient of the energy with respect to the domain and parameters. Experimental results on denoising and image labeling problems show that learning with truncated optimization greatly reduces computational expense compared to “full” fitting.