Dendrogram distance: an evaluation metric for generative networks using hierarchical clustering

Generative modeling is a task that aims to estimate the generation process of a given source dataset. Models obtained as a result of this approach can be used to sample novel data points that follow the distribution of the source training set, allowing for different applications in machine learning. Performing generative modeling using neural networks has become very popular mainly because of the success of Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) and later with Diffusion models (Luo, 2022). The GAN framework relies on two different networks, a generator and a discriminator, that compete against their selves to perform the generative task, as shown in Figure 1. Figure 1: Diagram that illustrates the different components of GAN. The generator network G transforms a random input z into samples that should be realistic, while the discriminator network D tells apart which samples came from the training data.