GAN-Tree: An Incrementally Learned Hierarchical Generative Framework for Multi-Modal Data Distributions

Abstract

Despite the remarkable success of generative adversarial networks, their performance seems less impressive for diverse training sets, requiring learning of discontinuous mapping functions. Though multi-mode prior or multigenerator models have been proposed to alleviate this problem, such approaches may fail depending on the empirically chosen initial mode components. In contrast to such bottom-up approaches, we present GAN-Tree, which follows a hierarchical divisive strategy to address such discontinuous multi-modal data. Devoid of any assumption on the number of modes, GAN-Tree utilizes a novel modesplitting algorithm to effectively split the parent mode to semantically cohesive children modes, facilitating unsupervised clustering. Further, it also enables incremental addition of new data modes to an already trained GAN-Tree, by updating only a single branch of the tree structure. As compared to prior approaches, the proposed framework offers a higher degree of flexibility in choosing a large variety of mutually exclusive and exhaustive tree nodes called GAN-Set. Extensive experiments on synthetic and natural image datasets including ImageNet demonstrate the superiority of GAN-Tree against the prior state-of-the-art. * equal contribution

Multi-Label Classification in Anime Illustrations Based on Hierarchical Attribute Relationships

“…In this subsection, we will explain the hierarchical divisive clustering (HDC) details. Specifically, in HDC, we adopt the hierarchical GAN-Tree [ 15 ] mentioned in Section 2.3 , which is used to perform the clustering. A hierarchical GAN-Tree is a hierarchical feature representation that transforms the original feature-embedding into a full binary tree.…”

“…Although these approaches address the issues of overfitting and mode collapse, they are unable to meet the requirements of multi-label classification with a single prior and the capacity of a single generator transformation. Recently, researchers have introduced a tree structure, called the hierarchical GAN-Tree [ 15 ], to facilitate the clustering by a multi-generator mode. This method can be utilized together with the corresponding prior distribution to generate samples with the desired level of quality and diversity.…”

“…Specifically, for the HDC part, we obtain feature representations of the anime illustrations belonging to different categories by divisive clustering. Inspired by the previous study [ 15 ], we use a clustering algorithm based on multiple generative adversarial networks (GANs) organized in a binary tree structure, which can obtain more appropriate category representations on datasets with a variety of styles, just as with the anime illustration dataset. In the HLA part, following the pre-defined hierarchical label structure from the dataset, the obtained feature representations are organized by hierarchical label assignments to form a feature with rich hierarchical relationships between labels.…”

Top-Down Deep Clustering with Multi-Generator GANs

“…Two key differences of our work is that (i) we use a separate generator (not a discrete latent variable) to encode a cluster, enabling our method to discover clusters with more representation capacity, and that (ii) we obtain hierarchical clusters, unlike the previous methods. Kundu et al (2019) propose the GAN-Tree framework, which slightly resembles our approach, since it also involves a hierarchical structure of independent nodes containing GANs capable of generating samples related to different levels of a similarity hierarchy. There are several differences, however.…”

Unsupervised Image Generation with Infinite Generative Adversarial Networks

“…where sampling from the posterior p(φ|x i , C) (previously p(φ|x i , Φ)) in Equation 10becomes feasible. Note our formulation is different from traditional CGANs and GANs with multiple generators [17,53,28,33] in: (1) our approach is still Bayesian. ( 2) we still model an infinite number of GANs and do not rely or impose assumptions on the prior knowledge of cluster numbers.…”