Fd-vae: A feature driven vae architecture for flexible synthetic data generation

Variational autoencoders (VAEs) are artificial neural networks used to learn effective data encodings in an unsupervised manner. Each input x provided to a VAE is indeed mapped to an internal representation, say z, in a low-dimensional space, called the latent space, from which an approximate version $$ ilde{x}$$ of x can be eventually reconstructed via a decoding phase. VAEs are very popular generative models because, by randomly sampling points from the latent space, we can generate novel and unseen data that still reflect the characteristics of the dataset used for the training. In many application domains, however, generating random instances is not enough. Rather, we would like mechanisms that can generate instances enjoying some high-level features that are desired by the users. To accomplish this goal, a novel VAE architecture—named Feature Driven VAE—is presented. Internally, it uses Gaussian Mixture Models to structure the latent space into meaningful partitions, and it allows us to generate data with any desired combination of features, even when that specific combination has been never seen in the training examples. The architecture is orthogonal to the underlying application domain. However, to show its practical effectiveness, a specialization to the case of image generation has been presented and implemented. Results of experimental activity conducted on top of it are eventually discussed.