We present the development of a portable embedded system for real-time electrocardiogram (ECG) beat classification for ventricular dysfunction related arrhythmias and conditions. These conditions are clinically relevant for monitoring heart failure patients and possible Left Ventricular Assist Device (LVAD) need, as well as potentially enabling the monitoring of remodeling of the native heart after LVAD implantation. We create a dataset combining single-lead standard datasets from Physionet repository. The system processes single-lead ECG signals using a lightweight, quantized, hardware-accelerated model based on residual connections. We run the proof of concept in an ARM Cortex-M33-based SoC, measuring an inference time of 119 milliseconds with 0.95 mJ of energy consumption at 78 MHz, Flash usage of 709.6KB, and runtime RAM usage of 24.7 KB, with an average AUC of 0.9946 in the multiclass classification task. We validated the system end-to-end using a patient simulator and the MAX30003 sensor. The proposed system enables continuous monitoring without external computation, paving the way towards integration with LVAD systems to allow for patient monitoring.
Segmentation of magnetic resonance images (MRI) facilitates analysis of human brain development by delineating anatomical structures. However, in infants and young children, accurate segmentation is challenging due to development and imaging constraints. Pediatric brain MRI is notoriously difficult to acquire, with inconsistent availability of imaging modalities, substantial non-head anatomy in the field of view, and frequent motion artifacts. This has led to specialized segmentation models that are often limited to specific image types or narrow age groups, or that are fragile for more variable images such as those acquired clinically. We address this method fragmentation with BabySeg, a deep learning brain segmentation framework for infants and young children that supports diverse MRI protocols, including repeat scans and image types unavailable during training. Our approach builds on recent domain randomization techniques, which synthesize training images far beyond realistic bounds to promote dataset shift invariance. We also describe a mechanism that enables models to flexibly pool and interact features from any number of input scans. We demonstrate state-of-the-art performance that matches or exceeds the accuracy of several existing methods for various age cohorts and input configurations using a single model, in a fraction of the runtime required by many existing tools.
This paper proposes a data-driven generative framework for learning ray-tracing behavior directly from base station and user equipment locations. The channel is represented in the angular-delay domain, and a conditional denoising diffusion probabilistic model (cDDPM) is trained to learn its distribution conditioned on spatial coordinates. To enhance the use of location information, the model employs a cross-attention mechanism. The proposed approach captures geometry-dependent channel behavior without requiring explicit 3D scene information. Experiments in a realistic urban environment show that it outperforms both a convolutional neural network baseline and a naive cDDPM in predicting dominant path characteristics. These results highlight the promise of cross-attention-enhanced diffusion models for location-conditioned and data-driven synthesis of ray-traced wireless channels.
Despite significant advancements in deep learning-based CSI compression, some key limitations remain unaddressed. Current approaches predominantly treat CSI compression as a source coding problem, neglecting transmission errors. In finite block length regimes, separate source and channel coding proves suboptimal, with reconstruction performance deteriorating significantly under challenging channel conditions. While autoencoder-based compression schemes can be adapted for joint source-channel coding, they struggle to accurately model increasingly complex channel distributions.We propose Residual-Diffusion Joint Source-Channel Coding (RD-JSCC), a novel framework that leverages diffusion models to learn robust CSI representations. Our architecture integrates a lightweight autoencoder with a residual diffusion module that iteratively refines CSI reconstruction, enabling: (i) graceful performance degradation across variable SNR conditions, and (ii) robust CSI estimation under realistic multipath fading scenarios. Our flexible decoding strategy balances computational efficiency and performance by dynamically selecting between low-complexity autoencoder decoding and sophisticated diffusion-based refinement based on channel conditions. Comprehensive simulations demonstrate that RD-JSCC significantly outperforms existing autoencoder-based approaches, particularly in challenging wireless environments. Furthermore, we demonstrate substantial decoding latency reduction through a single-step inference variant of our diffusion model, offering an efficient and practical solution for next-generation wireless communication systems.
Hybrid parasitic arrays are one approach to scale array aperture sizes with limited number of radio frequency (RF) chains and low power consumption. In this architecture, several active elements are connected to RF chains while parasitic elements are connected to digitally controlled reactances. Achieving good communication performance requires sufficient mutual coupling between active and parasitic elements. In this paper, we optimize the assignment of parasitic and active elements at the transmitter based on the signal-to-noise ratio (SNR) at the receiver. For a fixed fraction of parasitic antenna elements, this problem becomes a constrained assignment task, which is NP-hard due to its combinatorial nature. We employ a simulated annealing approach to find near-optimal antenna configurations that work well in a variety of channel conditions. The numerical results show how the optimized assignment patterns become beneficial for channels with stronger line-of-sight (LOS) component.
Passive correlation-based methods have revolutionized seismic imaging and temporal monitoring over the last several decades by implementing blind use of the complete source environment. Under the support of DARPA’s Defense Application of Innovative Remote Sensing program, we here present an initial study aimed at practical interpretation of array-based passive correlation radar methods with an emphasis on geometric untangling of range-Doppler images. Both isotropic and azimuthally biased source distributions are considered, and we describe through both real and synthetic examples how target information may be recovered in each scenario. We present results from single closely spaced linear arrays and show that interpretability, correlation function convergence, and an understanding of the source environment can lead to usable information. This precludes the need for the transmitter location and waveform modeling involved in classic passive radar approaches, and opens the possibility of using distributed receiver networks to broadly image and monitor environmental and anthropogenic targets.
Physical layer security (PLS) can be used to pro-vide keyless and innately secure communications by exploiting the propagation characteristics of the wireless channel at the physical layer (PHY) layer to co-generate cipher keys for symmetric encryption. While existing methods use channel state information as input for secret key generation, we consider a novel approach that derives features from frequency-time domain representation of in-phase and quadrature (IQ) samples to derive initial keys with high agreement rate. A recurrent neural network (RNN) with a quadruplet loss function is trained based on spectrograms collected from bi-directional time-division duplex channel probing between the two parties and on spectrograms collected from a passive eavesdropper. Information reconciliation is achieved using Reed Solomon codes. At the last step of the protocol, a 512-bit cryptographic key is generated by feeding the reconciled bit sequence to the Secure Hash Algorithm 3. Our experiments with software-defined radios in both indoor and outdoor environments evaluate the proposed key generation system under different channel propagation environments and hardware inaccuracies. We demonstrate that the proposed RNN-assisted channel feature extractor trained on indoor channel probe data operates successfully in two outdoor setups and produces up to 70 amplified keys out of 100 probes exchanged between two trusted parties i.e., Alice and Bob, while the passive eavesdropper fails to directly predict any of the generated keys.
Diffusion magnetic resonance imaging (dMRI) is widely used to map structural connectivity in the human brain, requiring the acquisition of similar to 100 diffusion-weighted images across multiple diffusion strengths (b-values) and gradient directions. Lower baseline SNR and higher acceleration factors remain a challenge for higher resolution dMRI, necessitating artificial intelligence (AI) based computational imaging methods to solve an inverse problem. However, existing AI methods yield limited gains in dMRI due to large SNR variations across b-values and phase inconsistencies introduced by diffusion gradients. In this work, we address these challenges with a novel loss-augmentation strategy for training physics-driven AI reconstructions. Our method boosts the SNR of high-b-value images through a combination of averaging and denoising. Phase adaptation is used in the former to handle phase mismatches among images, while a score-based approach is used for the latter. Experiments comparing our method with conventional reconstructions and standard AI methods demonstrate improved reconstruction quality.
In recent years, distributed wireless communication optimization, where training data is stored remotely from local multi-access edge computing (MEC) processors to preserve data security privacy and minimize complexity, has seen noteworthy progress for relevant wireless Internet-of-Everything (WIoE) networks beyond 6G. Nonetheless, the exploding number of WIoE clients requires secure data storage and scaled data processing at the network and transmitter, which local processors might be unable to afford. Parallel to this, we are witnessing widespread quantum-enabled learning adoptions for optimizing wireless communications. The rapid growth of quantum technologies has introduced security concerns for classical channels, due to their potential to undermine classical cryptographic approaches. This paper, therefore, considers the adoption of quantum-enabled learning with quantum communication protocol, especially quantum secure direct communication (QSDC) via superdense coding. While the processing learning happens across different locations for next-generation WIoE networks, the QSDC prevents vulnerabilities of data poisoning and model stealing in connected quantum collaborative learning.
Signal detection in environments with unknown signal bandwidth and time intervals is a basic problem in adversarial and spectrum-sharing scenarios. This paper addresses the problem of detecting signals occupying unknown degrees of freedom from non-coherent power measurements where the signal is constrained to an interval in one dimension or a hypercube in multiple dimensions. A Generalized Likelihood Ratio Test (GLRT) is derived, resulting in a straightforward metric involving normalized average signal energy on each candidate signal set. To overcome the inherent computational complexity of exhaustive searches, we propose a computationally efficient binary search method, reducing the complexity from O(N2) to O(N) for the one-dimensional case. Simulations indicate that the method maintains performance near exhaustive searches and achieves asymptotic consistency, with the interval-of-overlap converging to one under constant SNR as measurement size increases. The simulation studies also demonstrate superior performance and reduced complexity compared to contemporary neural network-based approaches, specifically outperforming custom-trained U-Net models in spectrum detection tasks.
We study online reinforcement learning (RL) with environmental safety constraints for a mobile robot navigating a maze with obstacles and unsafe zones, where the agent must maximize task return while respecting safety budgets. We model the problem as a Constrained Markov Decision Process (CMDP) and propose the Augmented Proximal Policy Optimization (APPO) for policy updates. A variational autoencoder (VAE) estimates from data the probability that a location is unsafe, which yields a calibrated cost signal to regularize the reward. Experiments indicate that APPO enables reliable navigation within safety-constrained regions while maintaining competitive returns.
Given multiple observations, the Wilks’ Lambda test can be used to detect a signal in interference-plus-noise whose covariance matrix is not known. Although signals to be detected in applications such a space-time adaptive processing are generally rank 1, losses in detection probability caused by uncertainties in the azimuth angle of a potential target within the transmitted beam and/or Doppler frequency can be mitigated by implementing a subspace version of the Wilks’ Lambda test. The subspace test increases the dimensionality of the hypothesized signal space beyond 1. Using multiple observations in a hypothesis test is equivalent to non-coherent integration and is expected to increase the probability of detection of any signal, including signals that are mismatched with the model. It is observed however that mismatched signal are rejected with higher probability by the subspace based Wilks’ Lambda test as the number of observations increase. This anomalous result is analytically explained in this paper.
This paper introduces a novel digital post-distortion (DPoD) technique designed to improve the received signal quality in multi-user orthogonal frequency-division multiple access (OFDMA) systems under severe power amplifier (PA) nonlinearities in the involved transmitters. The proposed approach combines memory polynomial-based nonlinear distortion modeling and cancellation with practical parameter estimation using demodulation reference signals (DMRS). Importantly, a new extended DMRS structure and corresponding receiver signal processing solutions are proposed which enable the receiver to mitigate both the ordinary pass-band nonlinearities as well as nonlinear crosstalk or interference between multiple frequency-multiplexed transmitters operating simultaneously within the channel bandwidth. Extensive numerical evaluations conducted in a 5G NR uplink scenario at 28GHz demonstrate significant receiver performance improvements, measured through effective error vector magnitudes (EVMs) for the frequency-multiplexed waveforms. The results validate the method’s effectiveness even under aggressive PA nonlinearities and high-order modulation schemes including 64-QAM and 256-QAM, thus paving the way towards improved power-efficiency and coverage in future OFDMA networks.
A graph neural network (GNN)-based method is proposed for radio map construction and uncertainty prediction in urban environment. Based on sparse measurements, the received signal strength (RSS) values are interpolated over the entire area. A graph is constructed using the city map and base station (BS) locations to capture the spatial correlation pattern of the RSS. Our proposed graph construction rule allows faithful reconstruction of the sharp transitions in the RSS values across line-of-sight (LOS) and non-line-of-sight (NLOS) regions. The uncertainty levels associated with the RSS estimates are jointly predicted using the negative log-likelihood (NLL) training cost. Numerical tests show that the proposed algorithm achieves up to a 30% reduction in root mean square error (RMSE) compared with an existing GNN-based approach, and significantly outperforms a convolutional neural network (CNN)-based method with a comparable parameter count. Furthermore, the uncertainty estimates are readily interpretable, typically assigning higher uncertainty in areas with larger prediction errors or fewer available samples.
This paper investigates a full-duplex (FD) multiple-input multiple-output (MIMO) setting for integrated sensing and communication (ISAC), which enables simultaneous monostatic sensing and communication with a single base station (BS). We consider frequency bands that exhibit sparse propagation characteristics, such as mmWave bands. A key challenge for these systems is the design of hybrid precoders and combiners that facilitate ISAC functionalities while mitigating self-interference (SI) that is caused by concurrent transmission and reception. While prior work has focused on hybrid precoder and combiner design for FD ISAC with prior communication channel and target parameter knowledge, the problem of joint channel and target parameter estimation remains unexplored. We address this gap by designing SI-aware hybrid training precoders and combiners that form a beam codebook optimized for sparse channel estimation. Our design minimizes the mutual coherence, a key metric in compressed sensing, while effectively suppressing the SI to enable accurate joint estimation. We evaluate our approach in terms of estimation accuracy, as well as SI mitigation performance.
Many imaging systems require accurate characterization of their forward operators for reliable reconstruction. While shift-invariant systems admit efficient convolutional representations, many practical imaging systems are shift-varying and cannot be captured by classical convolution models. We propose the Shift-Varying Neural Operator, an efficient and expressive architecture for learning spatially varying linear operators directly from measurements. Our method builds on existing factorizations of spatially varying convolutions and expresses them in a learnable architecture. Each layer implements a spatially adaptive transformation constructed from a low-rank factorization of modulated convolutional bases. Our experiments show that our proposed method accurately recovers spatially varying point spread functions (PSFs) and learns interpretable operators. Furthermore, we show that the learned forward operator can be integrated into existing iterative inverse problem solvers.
We propose a novel method for estimating intercarrier interference (ICI) matrices for orthogonal frequency division multiplexing (OFDM) on doubly-dispersive channels using the Karhunen-Loeve transform (KLT). While previous work has described KLT-based and basis expansion model (BEM) estimation of time-domain channel impulse responses, this work derives the relationship between the autocorrelation of the time-domain channel and that of its corresponding frequency-domain ICI matrix. We demonstrate that when the time-domain channel has a finite number of independent lags, the diagonals of the ICI matrix are stationary, resulting in a Hermitian autocorrelation structure, even if the underlying time-domain channel columns are non-stationary. For the considered time-varying Jakes wide-sensestationary-uniform-scattering (WSSUS) channel, we show that the rank of the frequency-domain autocorrelation is determined solely by the channel length. This yields substantial complexity reductions compared to time-domain estimation, particularly at the high Doppler spreads considered in this work. We integrate this estimator into a turbo-equalization scheme that iterates between a maximum a posteriori (MAP) equalizer and a channel decoder to jointly improve channel estimation and bit error rate (BER). Simulation results show the proposed method yields superior BER compared to orthogonal matching pursuit (OMP) and improved mean squared error (MSE) over standard timedomain KLT estimation techniques.
We propose a min-max uplink over-the-air (OTA) aggregation approach for training multiple models simultaneously through federated learning (FL) over wireless channels that vary across communication rounds. We design device grouping and joint transmit-receive beamforming to minimize the maximum expected optimality gap across models, ensuring min-max fairness. Through the OTA aggregation process, we establish an upper bound on the optimality gap and show that the T -round joint optimization problem can be broken down into individual per-round min-max design problems. We propose an efficient approach for the per-round problem by employing clustering-based device grouping and closed-form updates for beamforming optimization. Simulation results show that our proposed aggregation approach significantly outperforms both the traditional single-model training approach and other multi-model training methods.
Video quality assessment (VQA) is a key component in video processing and analysis, particularly with the growing volume of user-generated content (UGC) uploaded to platforms like YouTube and social media. In this work, we present a lightweight No-Reference Video Quality Assessment (NR-VQA), which combines a Video Masked Autoencoder (VMAE) and a modified Multilayer Perceptron Mixer (MixVPR) for quality prediction. The proposed VMAE and MLP-Mixer are based on a combination of a regressor inspired by the MixVPR feature mixer. The experimental results show that the proposed method achieves an excellent trade-off between performance and complexity, demonstrating its suitability for real-world applications. The algorithm’s performance is tested across five UGC datasets: KoNViD-1k and LIVE VQC, LIVE QUalcomm, CVD2014, and YouTube UGC. This work builds on advances in deep learning, particularly in transformer-based architectures and feature-mixing techniques.
We propose a time-frequency modulation and coding scheme for wireless communication systems in the presence of Linear Frequency Modulated (LFM) chirp radar interference that uses joint erasure-error decoding of Reed-Solomon codes to increase the codeword information rate of the signaling scheme compared to standard error-correcting decoding. The time-frequency modulation framework used takes advantage of the structure of LFM chirp waveforms in time-frequency to determine the positions of erasures in time-frequency resulting from LFM chirp interference. The resulting modulation scheme results in a relatively small number of interference-induced erasures, which can be corrected with joint erasure-error decoding of a Reed-Solomon code using a modified Berlekamp-Massey algorithm. The resulting performance shows that LFM chirp interference can be effectively mitigated with a relatively small reduction in overall code rate, and code rate can be further increased using joint erasure-error decoding when compared to standard error-correcting decoding.
