Automatic Correction by Deep Generative Model and its Application to Building Construction

Convolutional neural network has been successfully used for the recognition and generation of the data under a given distribution. The present work uses a deep generative model for a data correction, based on generative adversarial networks (GAN). After a generator is trained by a given data distribution, an encoder is trained to infer the obtained latent space of the generator to correct an input data. That is, if an input to the encoder is within the given data distribution, encoder is just an inverse mapping to generate the same output. On the other hand, if an input is slightly out of the distribution, then the encoder modifies the input to make the output within the distribution. In addition, a classifier is used for the fine-tuning of the data generation. The proposed model for data correction by the generator with encoder and classifier (ECGAN) is evaluated by computer experiments on a building design, where three-dimensional data should satisfy multiple physical constraints of resistance strength criteria.