Abstract: The present disclosure is directed to a system for reparameterizing of a neural network to optimize structural designs. The system can obtain data descriptive of a design space for a physical design problem. The design space is parameterized by a first set of parameters. The system can reparameterize the design space with a machine-learned model that comprises a second set of parameters. For a plurality of iterations, the system can provide an input to the machine-learned model to produce a proposed solution. The system can apply one or more design constraints to the solution to create a constrained solution. The system can generate a physical outcome associated with the constrained solution using a physical model. The system can evaluate the physical outcome using an objective function and update at least one of the second set of parameters. After the plurality of iterations, the system can output a solution.

Abstract: An invention based on learning a discrete recurrent neural network for a given signal domain is described. In one implementation to the domain of visual images, the method can be used to efficiently compress digital photographs and to devise a new perceptual distortion measure between images that well-matches data collected from a human psychophysics experiment. Other applications of the invention include unsupervised detection of recurrent patterns in high-dimensional data and Shannon-optimal error-correcting coding from few training examples.