Topic modeling plays a critical role in organizing and understanding large-scale web content. While neural topic models (NTMs) based on variational autoencoders (VAEs) have achieved notable success in analyzing textual data, they remain limited in addressing the multimodal nature of modern web content. Existing unimodal or multimodal extensions often suffer from posterior collapse and fail to capture the directional semantics inherent in both text and images, resulting in incoherent topics and limited interpretability. To address these challenges, we propose MM-vNTM (MultiModal Neural Topic Model with von Mises-Fisher Mixtures), a framework for web-scale topic discovery over multimodal data. MM-vNTM leverages pre-aligned cross-modal embeddings as inputs and jointly models document-level representations of text and image modalities in a shared hyperspherical latent space. Furthermore, it defines topics as mixtures of von Mises-Fisher (vMF) distributions in the L2-normalized word embedding space, explicitly capturing directional similarity. Experiments on multimedia web datasets demonstrate that MM-vNTM consistently outperforms state-of-the-art unimodal and multimodal baselines in terms of overall topic quality, highlighting its effectiveness for real-world web scenarios.

In recent years, axial data, where observations are treated as axes of direction, has gained prominence in a range of complex tasks, including gene expression data clustering, blind speech separation, and depth image analysis. However, prevailing methods for axial data modeling mainly rely on shallow probabilistic models, which often overlook the hidden and hierarchical dependencies in the latent space. These methods also require a separate, human-engineered feature extractor to obtain features from raw axial data for downstream tasks. This work introduces a novel framework, Axial Variational Autoencoders (AVAEs), for modeling and representation learning of axial data by leveraging a deep generative model, the Variational Autoencoder (VAE). Unlike existing approaches, our method can autonomously learn more expressive representations from axial data by designing a VAE that uses the Watson distribution as the latent prior. Furthermore, we introduce a tailored reparameterization technique to support stable training. We validate the effectiveness of our model through experiments on simulated axial datasets and a real-world application.

Clustering high-dimensional directional data (i.e., L2 normalized vectors) presents significant challenges due to the intricate spherical representations of latent embeddings and the limitations of classical (non-deep) clustering techniques. Moreover, dynamically inferring the number of clusters remains a fundamental issue in existing deep clustering methods, especially those involving complex model-selection criteria. This paper addresses these challenges by introducing a novel deep nonparametric clustering framework that employs hyperspherical latent embeddings within a Variational Autoencoder architecture, enhanced by an infinite Von Mises-Fisher Mixture Model as a dynamic prior. This approach enables automatic adaptation of cluster numbers during training, eliminating the need for predefined clusters and traditional model selection processes. Our scalable architecture effectively integrates In-vMFMM with hyperspherical embeddings to tackle the complexities of directional data. Utilizing a joint training strategy, our method alternates between updating neural network parameters and adjusting mixture model priors via nonparametric variational Bayes. Empirical evaluations on benchmark datasets, including complex ImageNet-50, demonstrate that our approach significantly outperforms state-of-the-art deep nonparametric clustering methods. It also robustly estimates the number of clusters, showcasing its effectiveness and versatility in handling high-dimensional directional data.

In this paper, we propose a clustering-based online news topic detection and tracking (TDT) approach based on hierarchical Bayesian nonparametric framework that allows topics to be shared across different news stories in a corpus. Our approach is formulated using the hierarchical Pitman-Yor process mixture model with the inverted Beta-Liouville (IBL) distribution as its component density, which has shown superior performance in modeling text data than the widely used Gaussian distribution. Moreover, we theoretically develop a convergence-guaranteed online learning algorithm that can effectively learn the proposed TDT model from a stream of news stories based on varational Bayes. The merits of our TDT approach are illustrated by comparing it with other well-defined clustering-based TDT approaches on different news data sets.

Fine-grained Image-text retrieval is challenging but vital technology in the field of multimedia analysis. Existing methods mainly focus on learning the common embedding space of images (or patches) and sentences (or words), whereby their mapping features in such embedding space can be directly measured. Nevertheless, most existing image-text retrieval works rarely consider the shared semantic concepts that potentially correlated the heterogeneous modalities, which can enhance the discriminative power of learning such embedding space. Toward this end, we propose a Cross-Graph Attention model (CGAM) to explicitly learn the shared semantic concepts, which can be well utilized to guide the feature learning process of each modality and promote the common embedding learning. More specifically, we build semantic-embedded graph for each modality, and smooth the discrepancy between two modalities via cross-graph attention model to obtain shared semantic-enhanced features. Meanwhile, we reconstruct image and text features via the shared semantic concepts and original embedding representations, and leverage multi-head mechanism for similarity calculation. Accordingly, the semantic-enhanced cross-modal embedding between image and text is discriminatively obtained to benefit the fine-grained retrieval with high retrieval performance. Extensive experiments evaluated on benchmark datasets show the performance improvements in comparison with state-of-the-arts.

In modern clinical medicine, electrocardiogram (ECG) is a common diagnosis technique of cardiovascular diseases. The purpose of this paper is to propose a novel model-based clustering approach for analyzing ECG data. Our approach is composed of two modules: representation learning and ECG data clustering. In the module of representation learning, a deep generative model referred to as the hyperspherical variational recurrent autoencoder (HVRAE) is developed to extract the representation of observed ECG data, based on the variational autoencoder (VAE) with long short-term memory (LSTM) networks. In the module of ECG data clustering, we develop a nonparametric hidden Markov model (NHMM) based on Dirichlet process in which the number of hidden states is inferred automatically during the learning process. Moreover, the emission density of each hidden state of our NHMM follows a mixture of von Mises-Fisher (VMF) distributions which have better capability for modeling ECG representations than other commonly used distributions (such as the Gaussian distribution). To learn the proposed VMF-based NHMM, we theoretically develop an effective learning algorithm based on variational Bayes. The merits of our model-based clustering approach for analyzing ECG data are verified through experiments on publicly available ECG data sets.

Many cognitive researches have shown the natural possibility of face-voice association, and such potential association has attracted much attention in biometric cross-modal retrieval domain. Nevertheless, the existing methods often fail to explicitly learn the common embeddings for challenging face-voice association tasks. In this paper, we present to learn discriminative joint embedding for face-voice association, which can seamlessly train the face subnetwork and voice subnetwork to learn their high-level semantic features, while correlating them to be compared directly and efficiently. Within the proposed approach, we introduce bi-directional ranking constraint, identity constraint and center constraint to learn the joint face-voice embedding, and adopt bi-directional training strategy to train the deep correlated face-voice model. Meanwhile, an online hard negative mining technique is utilized to discriminatively construct hard triplets in a mini-batch manner, featuring on speeding up the learning process. Accordingly, the proposed approach is adaptive to benefit various face-voice association tasks, including cross-modal verification, 1:2 matching, 1:N matching, and retrieval scenarios. Extensive experiments have shown its improved performances in comparison with the state-of-the-art ones.

Clustering has been a subject of extensive research in data mining, pattern recognition, and other areas for several decades. The main goal is to assign samples, which are typically non-Gaussian and expressed as points in high-dimensional feature spaces, to one of a number of clusters. It is well known that in such high-dimensional settings, the existence of irrelevant features generally compromises modeling capabilities. In this paper, we propose a variational inference framework for unsupervised non-Gaussian feature selection, in the context of finite generalized Dirichlet (GD) mixture-based clustering. Under the proposed principled variational framework, we simultaneously estimate, in a closed form, all the involved parameters and determine the complexity (i.e., both model an feature selection) of the GD mixture. Extensive simulations using synthetic data along with an analysis of real-world data and human action videos demonstrate that our variational approach achieves better results than comparable techniques.

In recent years, an increasing number of security threats have brought a serious risk to the internet and computer networks. Intrusion Detection System (IDS) plays a vital role in detecting various kinds of attacks. Developing adaptive and flexible oriented IDSs remains a challenging and demanding task due to the incessantly appearance of new types of attacks and sabotaging approaches. In this paper, we propose a novel unsupervised statistical approach for detecting network based attacks. In our approach, patterns of normal and intrusive activities are learned through finite generalized Dirichlet mixture models, in the context of Bayesian variational inference. Under the proposed variational framework, the parameters, the complexity of the mixture model, and the features saliency can be estimated simultaneously, in a closed-form. We evaluate the proposed approach using the popular KDD CUP 1999 data set. Experimental results show that this approach is able to detect many different types of intrusions accurately with a low false positive rate.