Topic modeling is a process that discovers key themes in unstructured text data by identifying the distribution of topics and words in a document, revealing hidden dimensions. Latent Dirichlet allocation is a widely used generative probabilistic topic model, but it cannot capture the dependency between topics. Generally, the topics within a document are primarily influenced by its overarching theme which naturally interrelates the topics. Thus, it is imperative to unveil such relationships between the topics. To this end, this study proposes a multilevel topic model (MTM) to unearth the hidden topic dependency in a corpus through multilevel latent structure. The MTM allows word-based topic proportions to vary across the higher-level latent structure. The parameters are estimated with a modified EM algorithm using an upward-downward approach to alleviate the computational complexity. Empirical studies on corpora have also been conducted on the multilevel topic model and the hierarchy of multilevel topic model have been interpreted. These analyses have demonstrated that the proposed multilevel topic model outperforms latent Dirichlet allocation in terms of systematic interpretability.

In recent years, the art landscape has undergone a considerable transformation with the emergence of AI-powered generative art tools, challenging traditional notions of artistic authenticity and ownership. The exponential growth of generative artwork sharing on social media platforms has created an urgent need to protect artists' intellectual properties from impersonation, forgery, and style appropriation. This study introduces an innovative, lightweight detection framework that efficiently distinguishes AI-generated art from human-created artwork by analyzing spatial domain features using tree-based ensembles. The study focuses on two prominent generative image architectures, StyleGAN2-ADA and Stable Diffusion, to explore the method's effectiveness across various classes of probabilistic deep generative models while incorporating JPEG compression considerations to reflect real-world social media conditions. The framework was evaluated across a diverse dataset of 10,000 images, achieving a detection accuracy of 94.43 % for StyleGAN2-ADA and 97.97 % for Stable Diffusion outputs on average across varying quality factors (QF). A key limitation observed is the lack of cross-architecture generalization-classifiers trained on one generative model do not reliably detect outputs from others, highlighting the need for architecture-agnostic detection strategies for real-world deployment. These results demonstrate comparable or better performance to existing deep learning solutions, requiring significantly less computational resources and training data. The proposed approach represents a significant step towards digital art authentication, offering a practical solution for real-time detection of AI-generated artwork in social media environments. Future work will focus on expanding the framework's capabilities to address emerging generative models and developing and integrating tools for automatic art authentication across various social media platforms.

MultiSP deciphers tissue structure and multicellular communication from spatial multi-omics data.

This work aims to explore the integration of novel and powerful deep learning techniques and intractable engineering problems, especially by adopting deep generative models to tackle model updating problems under uncertainty. A conditional denoising diffusion probabilistic model-based updating framework is presented to extend the field of deep generative models-based model updating methods. The diffusion model is a representative generative AI technique that employs a Markov chain to progressively add noise to data (forward process), then train a deep neural network to reverse this corruption (reverse process), enabling high-quality data generation. The conditional denoising diffusion extends the standard diffusion model, which guides data synthesis by injecting conditional inputs into the diffusion process. The conditional diffusion-based model updating framework consists of two primary neural networks: a conditional network and a denoising network. The conditional network can summarise the synthetic/measured response data into an informative fixed-length vector, called a conditional embedding, for guiding the training and denoising process of the denoising network. The denoising network can learn to predict the noise added in the forward process and denoise to generate the posterior samples conditioned on the conditional embedding. Both networks are trained jointly, and their architectures are flexible and problem-dependent. The framework is applied to solve a simulation-based problem, which is a customised version of the NASA and DNV Uncertainty Quantification Challenge 2025, and an experimental case study, which is a recently designed benchmark testbed with both experiment uncertainty and controllable parameter uncertainty.

Upon hearing objects collide, humans can estimate physical attributes such as material and mass. Although the physics of sound generation is well established, the inverse problem that listeners solve - of inferring physical parameters from sound - remains poorly understood. Classical accounts posit the use of acoustic cues that correlate with physical variables, but do not explain how humans might distinguish multiple concurrent physical causes. To study this problem, we built a probabilistic generative model of impact sounds, combining theoretical acoustics with statistics of object resonances measured from hundreds of everyday objects, and used it to synthesize and manipulate experimental stimuli. Humans accurately judged object properties from collision sounds. However, when both of the colliding objects varied, performance was impaired if the distribution of object resonances deviated from those measured in real-world objects. The results suggest that listeners use internal physical models to separate the acoustic contributions of objects in the world.

Understanding human behavior represents a paramount challenge in modern social systems. This task must be tackled with tools that both explain the mechanisms underlying the social dynamics and efficiently handle vast amounts of data. While agent-based models (ABMs) are generally used as simulating tools to describe social dynamics, their connection to data is lackluster. Instead, here we adopt a probabilistic machine learning approach for fitting ABMs to real data. To this end, we propose a variational inference (VI) framework that estimates the macroscopic and microscopic parameters of several opinion dynamics models. Our methodology encompasses three steps: (i) translation of the opinion dynamics models into probabilistic generative models (PGMs), (ii) relaxation of discrete variables to make the models differentiable, and (iii) estimation of the parameters and latent variables via stochastic VI (SVI). Experiments show that VI improves over existing methods in estimating discrete and continuous variables, both at the microscopic and macroscopic scales, in all four different categories of rules opinion dynamics models. Moreover, VI effectively estimates high-dimensional variables, up to 400 agent-level attributes, and is faster than the alternatives.

Predicting cellular responses to genetic perturbations is critical for advancing our understanding of gene regulation. While single-cell CRISPR perturbation assays such as Perturb-seq provide direct measurements of gene function, the scale of these experiments is limited by cost and feasibility. This motivates the development of computational approaches that can accurately infer responses to unmeasured perturbations from related experimental data. We introduce dbDiffusion, a generative framework that integrates diffusion models with classifier-free guidance derived from perturbation information, operating in latent space through a variational autoencoder. Diffusion models are probabilistic generative models that approximate data distributions by reversing a Markovian diffusion process, progressively denoising Gaussian noise into structured outputs. By exploiting biological similarities in gene expression profiles and relationships among perturbations, dbDiffusion enables the conditional generation of gene expressions for previously unobserved perturbations. In contrast to competing approaches, dbDiffusion does not rely on Large Language Model or foundation models, which have been found to yield unsatisfactory results. Rather, it leverages embeddings derived from measured perturbations to generalize to unseen perturbations, effectively transferring information across related experimental conditions. In benchmarking against state-of-the-art methods on Perturb-seq datasets, dbDiffusion demonstrates superior accuracy in predicting perturbation responses. A methodological innovation of dbDiffusion is the integration of prediction-powered inference, which corrects for biases inherent in generative models and enables statistically rigorous downstream tasks, including the identification of differentially expressed genes. By combining deep generative modeling with principled inference, dbDiffusion establishes a scalable computational framework for predicting and analyzing transcriptomic perturbation responses, significantly extending the utility of Perturb-seq experiments.

This research combines artificial intelligence generated content technology with style transfer methods to significantly improve the efficiency of generating and transmitting Yongju opera language content. To overcome fundamental challenges in traditional content creation, we first performed a comprehensive analysis of Yongju opera’s linguistic characteristics and established a multi-source dataset comprising 100 classical scripts, 50 contemporary scripts, and 500 lyric audio clips collected from the Ningbo Drama Research Institute and various opera platforms. The proposed framework features a dual-model architecture where the transformer-based conditional probabilistic generation (TFCPG) model acts as the content generator, while the conditional variational autoencoder (CVAE) functions as the style processor. The TFCPG transforms modern Chinese input into text complying with Yongju opera’s grammatical standards, and the CVAE enhances the text by incorporating dialect vocabulary and rhythmic patterns through style latent variables manipulation. Implemented in TensorFlow 1.4 with multi-task learning (batch size 32, Adam optimizer, learning rate 0.01), experimental results demonstrate that the TFCPG completes text generation in 33.32 min, representing a 56.6% reduction compared to the Transformer baseline. The model achieves a bilingual evaluation understudy (BLEU) score of 45.55 ± 1.32 (p < 0.01). In human evaluations, the system scored 4.35 for dialect authenticity, 4.18 for artistic expression, and 4.27 for cultural relevance. The CVAE component attained a BLEU score of 44.26 with 97.03% style transfer accuracy, exceeding the sequence to sequence (Seq2Seq) baseline by 10.28%. These comprehensive results confirm our approach’s effectiveness for Yongju opera language generation and style adaptation.

Detecting anomalies in multivariate time series (MTS) is a crucial task in areas like financial fraud detection and industrial equipment monitoring. Recent research has focused on developing unsupervised probabilistic models to identify anomalous patterns within MTS. However, many of these methods rely on fixed parameter mappings for each MTS, resulting in high computational costs and limited adaptability. To overcome these challenges, we introduce a novel Meta Variational Memory Transformer (MVMT). MVMT captures the diverse patterns across various MTS by encoding them into a set of memory units using a specially developed meta memory attention (MMA) module. Utilizing these learned memory units, we introduce a memory-guided probabilistic generative model that selects relevant memories as priors for latent states, resulting in more expressive MTS representations. A key feature of MVMT is that MMA provides a diversified prior in the latent space, ensuring the generation of various patterns. Finally, we implement a Transformer-based upward–downward variational inference process to estimate the posterior distribution of latent variables. Our extensive experiments on six datasets demonstrate the effectiveness of MVMT in one-for-all anomaly detection tasks.

Active Inference is a closed-loop computational theoretical basis for understanding behaviour, based on agents with internal probabilistic generative models that encode their beliefs about how hidden states in their environment cause their sensations. We review Active Inference and how it could be applied to model the human–computer interaction loop. Active Inference provides a coherent framework for managing generative models of humans, their environments, sensors and interface components. It informs off-line design and supports real-time, online adaptation. It provides model-based explanations for behaviours observed in HCI, and new conceptual tools with the potential to measure important concepts such as agency and engagement in interaction. We discuss how Active Inference offers a new basis for a theory of interaction in HCI, tools for design of modern, complex sensor-based systems, and integration of artificial intelligence technologies, enabling it to cope with diversity in human users and contexts. We discuss the practical challenges in implementing such Active Inference-based systems.

In this study, we address the challenge of analyzing worker behaviors in high‐mix, low‐volume production environments, where traditional supervised learning methods struggle owing to the lack of labeled data and task variability among workers. To overcome these issues, we propose a novel hierarchical approach for unsupervised behavior pattern extraction using the Gaussian process‐hidden semi‐Markov model and hidden semi‐Markov model. Unlike existing studies that focus on clustering human actions, our method segments complex motion data into meaningful action units and tasks, enabling a deeper understanding of worker behaviors. This two‐layer probabilistic generative model performs segmentation without pretraining on labeled datasets, which is advantageous in dynamic industrial contexts. Experiments using six‐dimensional time‐series wrist movement time‐series data from three workers engaged in assembly tasks show that our approach significantly improves segmentation accuracy compared with baseline methods. The results demonstrate its effectiveness in identifying distinct behavior patterns, highlighting the potential of our method to advance machine learning‐based work analysis in industrial environments. © 2025 The Author(s). IEEJ Transactions on Electrical and Electronic Engineering published by Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.

DGenNO: a novel physics-aware neural operator for solving forward and inverse PDE problems based on deep, generative probabilistic modeling

Abstract KRAS is a major oncogenic driver in pancreatic ductal adenocarcinoma (PDAC), mutationally activated in approximately 90% of cases. Mutations in this oncogene have been associated with more aggressive disease, poorer outcomes, and have remained hard to drug for over three decades. Recently developed small molecule inhibitors of KRAS (KRASi) have shown promising efficacy in advanced, previously treated PDAC patients but ultimately, most patients develop acquired resistance. Approximately half of these resistance cases do not present a putative genomic driver. In preclinical studies using genetically engineered mouse models of PDAC, we and others have demonstrated that cell state identity along the classical/epithelial-mesenchymal axis is a key determinant of response and resistance to KRASi. Notably, acute KRASi treatment initially induces a strong and selective bottlenecking of the malignant population in vivo, resulting in tumors that are predominantly classical with depleted mesenchymal characteristics. While these findings highlight a potent cell state-selective effect of acute KRASi treatment on PDAC cells, the underlying mechanisms driving the drug-induced cell state selection and plasticity remain poorly characterized. These mechanisms may reveal novel therapeutic targets for developing combination therapies that improve therapeutic responses. To comprehensively map and modulate epithelial-mesenchymal (E-M) cell state plasticity in response to KRASi treatment, we performed CRISPRi Perturb-seq (single-cell gene expression readout) with lineage tracing on a patient-derived PDAC cell line treated with the RAS(ON) multi-selective inhibitor RMC-7977. We captured 734,092 high-quality single-cell transcriptomic profiles and 156,119 unique clones across 60 genetic perturbations, including E- and M-specific transcription factors (TF) inferred to influence E-M plasticity from previous single-cell lineage tracing experiments. To optimally model clonal growth and transition rates, we profiled cells across a time series at day 6, day 15 and day 22, after 1 week of DMSO or KRASi treatment. We developed a hierarchical generative probabilistic model to jointly infer E↔M transition rates and cell state-specific growth rates, while capturing the effects of perturbations and KRASi treatment on state dynamics. Analysis of longitudinally tracked clones revealed that KRASi treatment specifically reduced the M-state growth rate while increasing the probability of M→E transitions compared to DMSO. Notably, we identified several TFs, displaying KRASi-dependent changes in state transition dynamics, either promoting (ELF3, MEIS2 among others) or decreasing (ZNF281, IRF9, among others) probabilities of M→E transitions when compared to non-targeting controls. These preclinical findings suggest that perturbation of these factors may be used as a paradigm to prevent cell state plasticity upon acute KRASi treatment and could be used to homogenize tumor populations towards a treatment-sensitive state. Citation Format: Julien Dilly, Mike Bogaev, Lynn Bi, Abigail Collins, Martin Jankowiak, Aziz Al'Khafaji, Kyle E. Evans, Mehrtash Babadi, Thouis R. Jones, Elisa Donnard, David T. Ting, Nir Hacohen, Dana Pe'er, Eric S. Lander, Andrew J. Aguirre, Arnav Mehta. Mapping and modulating epithelial-mesenchymal plasticity under RAS(ON) multi-selective inhibition in PDAC through lineage tracing and Perturb-seq [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Pancreatic Cancer Research—Emerging Science Driving Transformative Solutions; Boston, MA; 2025 Sep 28-Oct 1; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2025;85(18_Suppl_3):Abstract nr B120.

Predicting cellular responses to genetic perturbations is critical for advancing our understanding of gene regulation. While single-cell CRISPR perturbation assays such as Perturb-seq provide direct measurements of gene function, the scale of these experiments is limited by cost and feasibility. This motivates the development of computational approaches that can accurately infer responses to unmeasured perturbations from related experimental data. We introduce dbDiffusion, a generative framework that integrates diffusion models with classifier-free guidance derived from perturbation information, operating in latent space through a variational autoencoder (VAE). Diffusion models are probabilistic generative models that approximate data distributions by reversing a Markovian diffusion process, progressively denoising Gaussian noise into structured outputs. By exploiting biological similarities in gene expression profiles and relationships among perturbations, dbDiffusion enables the conditional generation of gene expressions for previously unobserved perturbations. In contrast to competing approaches, dbDiffusion does not rely on LLM or foundation models, which have been found to yield unsatisfactory results. Rather, it leverages embeddings derived from measured perturbations to generalize to unseen perturbations, effectively transferring information across related experimental conditions. In benchmarking against state-of-the-art methods on Perturb-seq datasets, dbDiffusion demonstrates superior accuracy in predicting perturbation responses. A methodological innovation of dbDiffusion is the integration of prediction-powered inference, which corrects for biases inherent in generative models and enables statistically rigorous downstream tasks, including identification of differentially expressed genes. By combining deep generative modeling with principled inference, dbDiffusion establishes a scalable computational framework for predicting and analyzing transcriptomic perturbation responses, significantly extending the utility of Perturb-seq experiments.

Generative probabilistic models have shown promise in designing artificial RNA and protein sequences but often suffer from high rates of false positives, where sequences predicted as functional fail experimental validation. To address this critical limitation, we explore the impact of reintegrating experimental feedback into the model design process. We propose a likelihood-based reintegration scheme, which we test through extensive computational experiments on both RNA and protein datasets, as well as through wet-lab experiments on the self-splicing ribozyme from the Group I intron RNA family where our approach demonstrates particular efficacy. We show that integrating recent experimental data enhances the model’s capacity of generating functional sequences (e.g. from 6.7% to 63.7% of active designs at 45 mutations). This feedback-driven approach thus provides a significant improvement in the design of biomolecular sequences by directly tackling the false-positive challenge.

Accelerated inverse design of broadband microperforated panel absorbers based on probabilistic generative model

Supervised learning remains the dominant paradigm for predictive modeling in data science, yet real-world deployments frequently fail due to fragile data pipelines, distributional shift, and optimistic evaluation. This article surveys supervised learning approaches with a focus on robustness—defined as the stability of predictive performance under perturbations to data, environment, or assumptions. We organize the model space into seven families: linear and generalized linear models; tree-based models; kernel methods; instance-based methods; probabilistic generative models; neural networks; and ensemble learning. For each family we discuss inductive biases, optimization, computational complexity, calibration, and typical failure modes. We then synthesize a method-agnostic workflow spanning dataset auditing, leakage prevention, feature engineering, resampling, hyperparameter tuning, model selection, and post-hoc reliability analysis (calibration, uncertainty, and drift monitoring). Robustness strategies—regularization, data augmentation, adversarial training, cost-sensitive learning, resampling for class imbalance, monotonic constraints, conformal prediction, and causal sensitivity analysis—are reviewed with practical guidance. Case vignettes from healthcare, finance, and operations illustrate trade-offs between accuracy, interpretability, and reliability. The paper concludes with open research directions, including integrating causal structure into supervised objectives, leveraging self-supervised pretraining for tabular data, distributionally robust optimization, and aligning evaluation with societal impact.

In extremely noisy communication scenarios, the bone-conducted microphone (BCM) speech codec is often combined with speech bandwidth extension to improve the BCM speech quality. However, this tandem approach leads to a complex system architecture. To address the problem, a scalable codec for BCM speech based on generative and diffusion probabilistic models is proposed in this paper. Specifically, a specialized codec architecture is constructed to encode BCM speech while complementing its high-frequency components. Then, a key feature extraction block is presented to address the diminishing memory capacity in shallow layers as the network depth increases. Next, considering the potential lack of high-frequency detail information, an overall refinement block is introduced to refine the reconstructed speech signals. Finally, based on the U-Net architecture, a diffusion probability model is proposed to upsample the input audio signal from a bandwidth of 8 kHz to a high-resolution audio signal with a bandwidth of 20 kHz and a sampling rate of 48 kHz. The proposed method can simultaneously encode and improve BCM speech quality using a single network. It supports different bitrate settings without architectural changes or retraining and dynamically adjusts transmitted data based on changing network load. Simulation experiments demonstrate its feasibility.

The integration of autonomous aerial vehicles (AAVs), cognitive radio (CR), and non-orthogonal multiple access (NOMA) presents a promising solution to significantly enhance the performance of future wireless networks. Achieving this integration requires cognitive self-awareness for intelligent resource allocation. In this paper, we address the problem of sum rate maximization in AAV-enabled cognitive NOMA uplink systems through the joint optimization of subchannel assignment and power allocation, while considering the AAV’s mobility. The traditional approach to finding the optimal solution requires an iterative or exhaustive search across all possible combinations of subchannel assignment, power allocation, and AAV position at each time slot, leading to excessive computational complexity. Furthermore, machine learning models, often trained on datasets that do not fully capture the complexity of real-world scenarios, struggle to handle non-stationary events effectively. To solve this nonconvex optimization challenge, we draw inspiration from active inference in cognitive neuroscience and propose a novel data-driven approach called the Active Generalized Dynamic Bayesian Network (Active-GDBN). The main idea is to process the unknown nonlinear input of an exhaustive search optimization algorithm using an Active-GDBN framework. This framework leverages a probabilistic generative model to learn the complex relationships and dependencies among subchannel assignments, power distributions, and the AAV’s mobility. The model is facilitated by continuous neuronal message passing in both discrete and continuous states to predict the optimal configuration. Numerical results show that the proposed approach achieves sum rate performance near the optimal exhaustive search and surpasses other baseline approaches.

The framing of water infrastructure in the news influences how the public perceives future infrastructure development and associated social-environmental risks. This study examines English-language newspaper coverage of the Ken-Betwa river link, the first component of India’s National River Linking Program (INRLP) to receive approval. Data for this analysis comprised 316 newspaper articles, collected via a keyword search in LexisNexis API, from seven Indian English-language newspapers (Free Press Journal (India), Hindustan Times, Indian Express, The Economic Times, The Hindu, The Times of India (TOI), and Times of India (Electronic Edition)) published between 2004 and 2022. By applying LDA topic modeling, a type of generative probabilistic model, to this dataset, this study examines how evolving media narratives frame water infrastructure in India. Our results identify 23 distinct topics and three dominant frames: (1) a government policy frame, (2) INRLP comparative frame, and (3) environmental conservation frame. We find that these frames evolve, with early coverage emphasizing feasibility and government-led negotiations, and later articles highlighting environmental risks. Our analysis shows how media discourse reflects institutional logic and infrastructure milestones. This study demonstrates the value of computational methods for longitudinal media analysis, has the potential to reveal shifts in public discourse, and highlights power dynamics in environmental reporting.