The growing demand for high-throughput and lowlatency wireless applications in 5 G and beyond, such as autonomous driving and extended reality, has increased the need for accurate beam prediction in millimeter-wave (mmWave) and subterahertz (sub-THz) systems. Machine learning offers a promising solution, but practical deployment is often limited by model complexity and data labeling requirements. This paper investigates two complementary techniques to address these challenges: knowledge distillation (KD) and active learning (AL). KD enables compact models to approximate the performance of larger models, while AL reduces labeling effort by selecting informative training samples. We evaluate both techniques using a vehicle-toinfrastructure scenario from the DeepSense 6G dataset. Results show that KD improves Top-1 beam prediction accuracy by up to 12%, while AL reduces labeled data requirements by up to 30%. These findings highlight the potential of KD and AL for enabling efficient, scalable machine learning solutions in next-generation wireless systems.
Diffusion models have emerged as a powerful image priors for inverse problems. However, their performance can degrade significantly when there is a distribution shift between training and testing datasets. Quantifying distribution shifts is thus essential for ensuring the reliability of diffusion model priors. Existing shift detection methods typically require access to clean data, which is often unavailable in inverse problems. We propose a measurement-domain KL divergence estimator for linear inverse problems, enabling shift quantification directly from corrupted measurements. We prove that the proposed metric can recover the KL divergence between training and test distributions using only noisy observations and pretrained diffusion models. Our method provides a practical, unsupervised approach for detecting distribution shifts in inverse problems such as image deblurring—without relying on clean images.
In this paper, we propose a reflection design solution for a single-user reconfigurable intelligent surface aided mmWave MIMO communication system with multiple streams, by formulating it as a spectral efficiency maximization problem. A heuristic low-complexity reflection design (LCRD) method is proposed in the form of a non-iterative algorithm. It is then compared to benchmark methods and is found to demonstrate an extremely low complexity rendering it highly efficient for practical applications with only a slight trade-off in accuracy as compared to more computationally intensive methods.
This paper introduces a novel approach designed to significantly reduce this cost by mapping the spatial Digital beamforming is a critical technology for next-generation satellite communications. Conventional solutions in current missions primarily rely on hybrid analog-digital architectures due to the significant power consumption challenges of a fully-digital implementation. For future systems, Fast Fourier Transform (FFT)-based beamforming has been proposed as an efficient digital alternative, especially for managing many beams. However, the use of a FFT beamforming alone is not compatible with non-uniform antennas and beams. This inflexibility is a major problem for modern satellite systems that need beam reconfigurability to meet changing traffic demands. The NonUniform FFT (NUFFT) offers the needed flexibility for nonuniform grids, but it involves higher computational costs due to extra processing steps. This paper presents a new method to lower this cost by mapping the spatial samples onto a hexagonal grid instead of the traditional Cartesian grid. Taking advantage of the efficient 2D sampling efficiency of the hexagonal lattice, our approach uses circularly symmetric windows more effectively. This efficiency allows for more accurate signal reconstruction with fewer operations. The hexagonal transformation offers a more power-efficient and adaptable way to achieve real-time, dynamic beam steering required in advanced satellite systems.
Signal folding is an inherent problem encountered in phase reconstruction and sampling hardware. The precision of analogue-to-digital converter (ADC) measurements depends on the available dynamic range and number of ADC bins. Recent developments in hardware design have taken advantage of signal continuity to obtain precise measurements of high dynamic range signals using low-cost ADCs with modulo-sampling. The existing reconstruction algorithms provide batch-based methods for point estimates of the original signal, up to an unknown additive constant. We present a Bayesian model using a state-space model (SSM) capable of running in an online setting. Modulosampling and ADC quantisation non-linearities are modelled in the particle filter likelihood. We provide bounds to adaptively select the required sampling range to account for modulosampling discontinuities. This model accurately reconstructs a signal measured on real hardware, and significantly out-performs existing methods on a noisy dataset.
Prediction of a system’s future states is an essential problem to many applications and often invites a probabilistic view of the likelihood or uncertainty associated with a potential outcome. Extending recent work in stochastic interpolants and hierarchical rectified flow, we propose a generative modeling approach for probabilistic forecasting by learning and sampling from the conditional distribution of the future system state given its current state. The training data are constructed by pairing various initial states with their corresponding j-step future states from sequential time series observations. We apply two-layered hierarchical rectified flow to learn an acceleration field that governs the generation of random velocity samples, which are used to map the current state to a probabilistic ensemble of j-step-ahead forecasts. We characterize the velocity distribution for linearly interpolated samples from coupled source and target data and prove that the proposed procedure generates samples from the target conditional distribution. In addition, we propose to finetune the hierarchical rectified flow model at a specific flow time to better capture the pertinent velocity distribution. We demonstrate the advantages of our approach for short-term forecasting of a lowdimensional jump-diffusion process and the Lorenz 96 model.
While real world data frequently comprises multiple distinct modalities, most generative AI operates primarily with a single modality. Moreover, existing multimodal generation works primarily tackle either text-image generation or conditional modality generation. We extend the latent diffusion framework into the multimodal image domain, facilitating consistent joint generation of two or more modalities. Based on the idea that two modalities comprise some shared information and some information unique to each modality, we develop a variational autoencoder architecture and training scheme to generate an effective latent space for diffusion generation. We demonstrate generation of both RGB & semantic segmentation and RGB & depth map modality pairs. We find our disentangled latent space structure to improve generative quality and ensure consistency between modalities.
An online Bayesian changepoint detection framework is proposed to identify structural changes in multiple time series. An existing approach is extended to support asynchronous, multi-source inputs via both deterministic and probabilistic fusion strategies. The resulting framework enables timely, interpretable, and sensor-agnostic detection of forest changes, addressing key limitations of traditional offline and singlesensor methods. Experiments are conducted using both synthetic data and real Sentinel-1 SAR and Sentinel-2 optical data over tropical forests affected by deforestation. Results highlight the benefits of multi-source fusion for accurate and timely disturbance detection.
Integrated Sensing and Communication (ISAC) is a key paradigm for improving resource efficiency in wireless networks by merging sensing and communication functionalities. Orthogonal Time Frequency Space (OTFS) modulation exhibits robustness in the high-Doppler environments, which are common to future high-mobility applications. In this paper, we propose a novel ISAC-OTFS system based on a monostatic multiple-input multiple-output (MIMO) radar configuration, which enables a flexible trade-off between communication and sensing. This tradeoff is achieved by reserving a small subset of time–frequency (TF) domain resources as private for specific transmit antennas. Signals received on these private bins are used to form a virtual array (VA), thereby enhancing angle resolution. However, the introduction of private bins reduces the number of Doppler–delay (DD) domain bins available for data communication, establishing the fundamental trade-off between sensing performance and communication rate. Our design incorporates a non-ideal rectangular pulse shaping function at both the transmitter and receiver, and reformulates our previous work—originally based on ideal, bi-orthogonal pulses—to account for the interference introduced in this practical scenario. Simulation results validate that the proposed framework achieves high sensing performance comparable to systems using ideal pulses, with only a small impact on communication performance.
We consider covariance estimation under Toeplitz structure. Numerous sophisticated optimization methods have been developed to maximize the Gaussian log-likelihood under Toeplitz constraints. In contrast, recent advances in deep learning demonstrate the surprising power of simple gradient descent (GD) applied to overparameterized models. Motivated by this trend, we revisit Toeplitz covariance estimation through the lens of overparameterized GD. We model the covariance as a sum of K complex sinusoids with learnable parameters and optimize them via GD. We show that when $K=P$ (the matrix dimension), GD may converge to suboptimal solutions. However, mild overparameterization ($K=2 P$ or $4 P$) consistently enables global convergence from random initializations. Our experiments demonstrate that overparameterized GD can match or exceed the accuracy of state-of-the-art methods in challenging settings, while remaining simple and scalable.
This paper proposes a scalable method for identifying interactions in higher-order networks from observations of nodal processes. Finding such dependencies is important in many disciplines, including neuroscience, social influence modeling, and beyond. However, current approaches are either limited to extracting pairwise dependencies or struggle with scalability, as estimating higher-order dependencies becomes computationally prohibitive. To overcome these challenges, we introduce a tensorbased graph Volterra model that leverages low-rank decomposition techniques to estimate higher-order interactions efficiently. Our approach not only reduces computational and storage complexity but also acts as an implicit regularizer, improving network estimation in ill-posed settings. We validate our method through simulations and real data experiments, demonstrating competitive performance and enhanced scalability compared to existing techniques.
Estimation of differences in conditional independence graphs (CIGs) of two time series Gaussian graphical models (TSGGMs) is investigated where the two TSGGMs are known to have similar structure. The TSGGM structure is encoded in the inverse power spectral density (IPSD) of the time series. In several existing works, one is interested in estimating the difference in two precision matrices to characterize underlying changes in conditional dependencies of two sets of data consisting of independent and identically distributed (i.i.d.) observations. In this paper we consider estimation of the difference in two IPSDs to characterize the underlying changes in conditional dependencies of two sets of time-dependent data. We extend an existing group lasso-penalized D-trace loss function approach in the frequency domain for differential time-series graph learning to non-convex group log-sum penalized D-trace loss function approach to improve performance. An alternating direction method of multipliers (ADMM) algorithm is presented to optimize the objective function. A synthetic data example is presented in support of the proposed approach where our proposed log-sum penalized loss significantly outperforms existing approaches with $F_{1}$ score as the performance metric.
Reversible Jump Markov Chain Monte Carlo (RJMCMC) methods are commonly used to estimate models involving discrete parameters, but are slow to converge, with high rejection rates for proposals. We embed RJMCMC proposals into a Sequential Monte Carlo (SMC) sampler, taking advantage of the inherent parallelism of SMC to decrease overall computation time. Additionally, we use the information encoded in the weights of particles in the SMC sampler to inform the probability of particular discrete proposals at each step. We demonstrate that these approaches offer comparable performance and improvements in computation time over conventional RJMCMC in the setting of estimating a finite mixture model with an unknown number of components.
Low-photon imaging is widely applied in many fields such as medical imaging, astronomy, and fluorescence microscopy. Compared to Poisson noise, the negative binomial (NB) distribution provides a more flexible and accurate model for over-dispersed photon-limited data. While NB-based reconstruction and denoising methods have demonstrated improved performance across various scenarios, accurately estimating the over-dispersion parameter r remains a significant challenge. Traditional approaches, such as the method of moments and maximum quasi-likelihood estimation, can struggle with accuracy and robustness. To address this limitation, we propose a deep learning-based framework to predict the over-dispersion parameter r in a data-driven manner. Experiments show that our method is both accurate and generalizable, enabling more effective applications of NB models in real-world imaging tasks.
Recent advances in forensic detection techniques, such as the Diffusion Noise Feature (DNF), have enabled highly accurate identification of AI -generated images by exploiting subtle statistical patterns introduced during the denoising process of diffusion models. In this work, we propose an antiforensic diffusion framework that regularizes the image generation process to produce images with DNF representations that closely resemble those of real images. Our approach integrates a DNF-guided regularization term into the reverse diffusion process, effectively minimizing the statistical distance between the generated and real image DNF distributions. Experimental results demonstrate that our method significantly reduces the detection accuracy of state-of-the-art forensic detectors, demonstrating a more realistic image generation.
A non-parametric method is introduced to estimate the measurement model of dynamical systems. The method uses a neural network trained in an unsupervised manner integrated into a particle filter framework. The network learns the measurement likelihood directly from the distribution of particles. The performance of the resulting neural network particle filter is first evaluated on synthetic data with a known measurement model showing a very interesting performance. The particle filter is then applied to sensor calibration with a specific focus on camera distortion estimation. Experimental results show that the method provides a reliable alternative to traditional parametric calibration techniques.
We propose a framework for time series prediction that leverages Koopman operator theory, a powerful tool for capturing nonlinear dynamical systems through linear evolution via transformation to a high dimensional space. By approximating the Koopman operator using a finite set of basis functions, we transform the forecasting problem into linearized propagation of observables. To account for parameter uncertainty, we adopt a Bayesian approach to infer the parameters of the approximate Koopman operator, enabling posterior sampling over dynamics. Our model incorporates prior knowledge via explicit parameter priors and performs inference by sampling from the posterior. The resulting Bayesian Koopman operator can be implemented with standard deep learning architectures such as linear recurrent neural networks. We demonstrate its advantages on complex dynamical systems, comparing performance to standard deterministic Koopman baselines.
Principal component analysis (PCA) and probabilistic (P) PCA offer linear models that account for low-dimensional structure of high-dimensional datasets. Deep generative models, on the other hand, such as variational autoencoders (VAEs), can capture nonlinear mappings, but often rely on a simple isotropic prior, which limits their ability to represent complex and clustered manifolds. This work develops a generative latent variable model that integrates a mixture-of-Gaussians (MoG) prior with low-rank constraints in the latent space, where each mixture component defines a local low-dimensional subspace. A critical contribution is that the MoG parameters, including both means and covariances, are learned directly from data. For unsupervised tasks, a sparse router network selects relevant ‘experts’ in the mixture-of-experts setting, thereby promoting disentangled and interpretable latent representations. Experiments on MNIST dataset demonstrate that the novel approach generates higher-quality, more consistent samples than standard VAEs or (P)PCA, and better aligns the latent manifold with the true data geometry. By combining nonlinear generation, low-rank covariance structure, and a sparse routing network, the proposed method reinforces deep generative modeling with principled manifold learning, while bridging the gap between supervised and unsupervised setups.
We present a new framework for representing and reconstructing multidimensional MR images via adaptive, feature-based representations. Specifically, we propose a disentangled representation model that separates different types of image features, such as geometry and contrast, into distinct lowdimensional latent spaces, reducing the degrees-of-freedom for multidimensional images and enabling flexible constraints on different latents/features. A latent diffusion model was introduced to capture the variations of these disentangled latents, providing feature-level generative priors for regularized reconstruction. New formulations and algorithms were developed to integrate the pretrained representation with task-specific adaptation for reconstruction from undersampled or noisy measurements, without task-specific training. We demonstrated effective disentanglement of geometry and contrast features in multi-contrast images, and its utility in different multidimensional MR applications, i.e., accelerated MR parameter mapping and self-supervised multicontrast denoising.
Fast Radio Bursts (FRBs) are brief, highly energetic astrophysical phenomena. Detecting FRBs is particularly challenging due to the prevalence of radio frequency interference (RFI), which often obscures or mimics genuine FRB signals, leading to a high rate of false positives. To address this issue, we propose a dual-headed architecture that leverages two learning paradigms: contrastive learning and a selective reconstruction head designed for open-set anomaly detection. Our architecture incorporates a modified contrastive loss to account for both the variability and the class imbalance inherent in transient detection. Furthermore, by confining the reconstruction process to normal data (consisting of RFI and noise), the proposed approach enables robust discrimination between FRBs and interference signals.
