In recent times, there has been a growing interest in the wireless community regarding the integration of intelligent reflecting surface (IRS) assistance in millimeter-wave (mmWave) and terahertz (THz) communications. This research paper sets out to develop a beam-centric multiple-access strategy tailored for this emerging paradigm. The fundamental concept involves employing multiple sub-arrays within a hybrid digital-analog array to create distinct beams. Each beam is then directed independently towards the desired direction, effectively mitigating inter-user interference and suppressing undesired signal reflections. The proposed approach amalgamates the benefits of both orthogonal multiple access (ensuring no inter-user interference) and non-orthogonal multiple access (utilizing the full time-frequency resource spectrum). As a result, this strategy has the potential to significantly enhance system capacity, a conclusion supported by Monte-Carlo simulations.

Every now and then, we witness significant improvements in the performance of Deep Learning models. A typical cycle of improvement involves enhanced accuracy followed by reduced computing time. As algorithms get better at their job, it is worthwhile to try to evaluate their performance on problems that are affected by them. Computationally intense problems, such as object detection for Computer Aided Laparoscopy (CAL), can benefit from such improvements in such technologies. Recently a new set of variants of You Look Only Once (YOLO) models based on Neural Architecture Search (NAS) technique have been released. Deci, the enterprise behind this new development, touts a much better performance both in terms of accuracy as well as computational efficiency. In this paper, we have analyzed the performance YOLO-NAS on a well-known benchmark dataset related to CAL. We found that the performance of all the NAS-based YOLO was inferior as compared to other State-of-the-Art (SoTA) YOLO models. We compare our results against the YOLOv7 model too.

In recent years, Sound AI is being increasingly used to predict machine failures. By attaching a microphone to the machine of interest, one can get real time data on machine behavior from the field. Traditionally, Convolutional Neural Net (CNN) architectures have been used to analyze spectrogram images generated from the sounds captured and predict if the machine is functioning as expected. CNN architectures seem to work well empirically even though they have biases like locality and parameter sharing which may not be completely relevant for spectrogram analysis. With the successful application of transformer-based models in the field of image processing starting with Vision Transformer (ViT) in 2020, there has been significant interest in leveraging these in the field of Sound AI. Since transformer-based architectures have significantly lower inductive biases, they are expected to perform better than CNNs at spectrogram analysis given enough data. This paper demonstrates the effectiveness of transformer-driven architectures in analyzing Sound data and compares the embeddings they generate with CNNs on the specific task of machine fault detection.

Bit and power loading (BPL) algorithms played a pivotal role in the success of orthogonal frequency division multiplexing (OFDM) in digital transmission, including lightemitting diode (LED) based wireless optical communications. Nevertheless, the conventional BPL algorithms do not distinguish the nonlinear distortion generated in LED transmitters from an additive noise, which leaves room for improvement. This letter presents a novel power loading and two BPL algorithms that maximize the transmission capacity while minimizing the nonlinear distortion generated in LED. The effectiveness of the proposed algorithms is evaluated through simulations and transmission experiments, revealing a throughput increase of up to 10% in comparison to what can be achieved employing classical algorithms.

Symmetric nonnegative matrix factorization (Sym-NMF) is an unsupervised method that seeks to approximate a symmetric matrix A, often a similarity or kernel matrix, by the product of low-rank factors. Most existing models deal with a factorization of the form A ≈ HH T , which may not yield a satisfactory approximation in cases where the matrix is indefinite. To address this limitation, this paper proposes a new SymNMF model based on the more general formulation A ≈ HSH T , that is more stable. To ensure a low-rank representation on the positive semi-definite cone, the LogDet divergence is incorporated as a regularization term. An alternative optimization algorithm is proposed to optimize the proposed objective. Experimental results on various synthetic and real-world datasets demonstrate that the proposed model outperforms or performs comparably to many state-of-the-art methods while maintaining superior computational efficiency.

In the field of dentistry, there is a growing demand for increased precision in diagnostic tools, with a specific focus on advanced imaging techniques such as computed tomography, cone beam computed tomography, magnetic resonance imaging, ultrasound, and traditional intra-oral periapical X-rays. Deep learning has emerged as a pivotal tool in this context, enabling the implementation of automated segmentation techniques crucial for extracting essential diagnostic data. This integration of cutting-edge technology addresses the urgent need for effective management of dental conditions, which, if left undetected, can have a significant impact on human health. The impressive track record of deep learning across various domains, including dentistry, underscores its potential to revolutionize early detection and treatment of oral health issues. Objective: Having demonstrated significant results in diagnosis and prediction, deep convolutional neural networks (CNNs) represent an emerging field of multidisciplinary research. The goals of this study were to provide a concise overview of the state of the art, standardize the current debate, and establish baselines for future research. Method: In this study, a systematic literature review is employed as a methodology to identify and select relevant studies that specifically investigate the deep learning technique for dental imaging analysis. This study elucidates the methodological approach, including the systematic collection of data, statistical analysis, and subsequent dissemination of outcomes. Results: In incorporating 45 studies, we identified selection criteria and research objectives, addressing significant gaps in the existing literature. These studies assist clinicians in examining dental conditions and classifying dental structures, including caries detection and the identification of various tooth types. We evaluated model performance, addressing the identified gaps, using diverse metrics that we strive to list and explain. Conclusion: This work demonstrates how Convolutional Neural Networks (CNNs) can be employed to analyze images, serving as effective tools for detecting dental pathologies. Although this research acknowledged some limitations, CNNs utilized for segmenting and categorizing teeth exhibited their highest level of performance overall.

Short-packet communication (SPC) and unmanned aerial vehicles (UAVs) are anticipated to play crucial roles in the development of 5G-and-beyond wireless networks and the Internet of Things (IoT). In this paper, we propose a secure SPC system, where a UAV serves as a mobile decode-and-forward (DF) relay, periodically receiving and relaying small data packets from a remote IoT device to its receiver in two hops with strict latency requirements, in the presence of an eavesdropper. This system requires careful optimization of important design parameters, such as the coding blocklengths of both hops, transmit powers, and the UAV’s trajectory. While the overall optimization problem is nonconvex, we tackle it by applying a block successive convex approximation (BSCA) approach to divide the original problem into three subproblems and solve them separately. Then, an overall iterative algorithm is proposed to obtain the final design with guaranteed convergence. Our proposed low-complexity algorithm incorporates robust trajectory design and resource management to optimize the effective average secrecy throughput of the communication system over the course of the UAV-relay’s mission. Simulation results demonstrate significant performance improvements compared to various benchmark schemes and provide useful design insights on the coding blocklengths and transmit powers along the trajectory of the UAV.

A document by Luis Hernandez Corporales . Click on the document to view its contents.

The Just Noticeable Difference (JND) refers to the smallest distortion in an image or video that can be perceived by Human Visual System (HVS), and is widely used in optimizing image/video compression. However, accurate JND modeling is very challenging due to its content dependence, and the complex nature of the HVS. Recent solutions train deep learning based JND prediction models, mainly based on a Quantization Parameter (QP) value, representing a single JND level, and train separate models to predict each JND level. We point out that a single QPdistance is insufficient to properly train a network with millions of parameters, for a complex content-dependent task. Inspired by recent advances in learned compression and multitask learning, we propose to address this problem by (1) learning to reconstruct the JND-quality frames, jointly with the QP prediction, and (2) jointly learning several JND levels to augment the learning performance. We propose a novel solution where first, an effective feature backbone is trained by learning to reconstruct JNDquality frames from the raw frames. Second, JND prediction models are trained based on features extracted from latent space (i.e., compressed domain), or reconstructed JND-quality frames. Third, a multi-JND model is designed, which jointly learns three JND levels, further reducing the prediction error. Extensive experimental results demonstrate that our multi-JND method outperforms the state-of-the-art and achieves an average JND1 prediction error of only 1.57 in QP, and 0.72 dB in PSNR. Moreover, the multitask learning approach, and compressed domain prediction facilitate lightweight inference by significantly reducing the complexity and the number of parameters.

This paper explores the realm of intelligent systems through an analogy inspired by RLC circuits, delving into the interconnected dynamics of reasoning, learning, and control. Leveraging the simplicity and clarity of the analogy, we navigate the conceptual landscape by drawing parallels between electrical components and the cognitive functions of modern AI. The presented analogical framework is the conclusion of the personal experiences of the author in developing intelligent systems, sparked by conversations with fellow researchers and students and presentations of research outcomes. It is worth recognizing the limitations of this analogy, as its reductionist nature may simplify the complexities inherent in intelligent systems. Yet, this exploration provides a fresh perspective on the foundational components of intelligent systems through the lens of the well-established RLC circuit theory.

Reconfigurable intelligent surfaces (RISs) operate similarly to electromagnetic (EM) mirrors and remarkably go beyond Snell law to generate an applicable EM environment allowing for flexible adaptation and fostering sustainability in terms of economic deployment and energy efficiency. However, the conventional RIS is controlled through high-latency field programmable gate array or micro-controller circuits usually implementing artificial neural networks (ANNs) for tuning the RIS phase array that have also very high energy requirements. Most importantly, conventional RIS are unable to function under realistic scenarios i.e, high-mobility/low-end user equipment (UE). In this paper, we benefit from the advanced computing power of neuromorphic processors and design a new type of RIS named NeuroRIS, to supporting high mobility UEs through real time adaptation to the ever-changing wireless channel conditions. To this end, the neuromorphic processing unit tunes all the RIS metaelements in the orders of ns for particular switching circuits e.g., varactors while exhibiting significantly low energy requirements since it is based on event-driven processing through spiking neural networks for accurate and efficient phase-shift vector design. Numerical results show that the NeuroRIS achieves very close rate performance to a conventional RIS-based on ANNs, while requiring significantly reduced energy consumption with the latter.

This paper proposes block-level precoder (BLP) designs for a multi-input single-output (MISO) system that performs joint sensing and communication across multiple cells and users. The Cramer-Rao-Bound for estimating a target's azimuth angle is determined for coordinated beamforming (CBF) and coordinated multi-point (CoMP) scenarios while considering inter-cell communication and sensing links. The formulated optimization problems to minimize the CRB and maximize the minimum-signal-to-interference-plus-noise-ratio (SINR) are nonconvex and are represented in the semidefinite relaxed (SDR) form to solve using an alternate optimization algorithm. The proposed solutions show improved performance compared to the baseline scenario that neglects the signal component from neighboring cells.

In this work, we perform a comparative analysis of discrete-and continuous-time estimators of information-theoretic measures quantifying the concept of memory utilization in short-term heart rate variability (HRV). Specifically, considering heartbeat intervals in discrete time we compute the measure of information storage (IS) and decompose it into immediate memory utilization (IMU) and longer memory utilization (MU) terms; considering the timings of heartbeats in continuous time we compute the measure of MU rate (MUR). All measures are computed through model-free approaches based on nearest neighbor entropy estimators applied to the HRV series of a group of 15 healthy subjects measured at rest and during postural stress. We find, moving from rest to stress, statistically significant increases of the IS and the IMU, as well as of the MUR. Our results suggest that both discrete-time and continuous-time approaches can detect the higher predictive capacity of HRV occurring with postural stress, and that such increased memory utilization is due to fast mechanisms likely related to sympathetic activation.

Affective patterns based on physiological signals reflect bodily changes linked to specific emotional states. Previous studies on the cardiac electrical signal, a key peripheral physiological signal, were limited by the measurement density of single-lead ECG signal, focusing solely on temporal pattern analysis but ignoring topographic pattern analysis that can reflect the body's emotional response. Our research advances affective peripheral pattern studies by innovatively using body surface potentials to comprehensively monitor cardiac electrical activity with increased measurement density. To tackle the challenge of extracting spatial and temporal features from multi-channel body surface potentials, we establish a dynamic correlation among these diverse channel signals through covariance matrices. Our hypothesis is that the dynamic inter-channel relationship provides a valuable source of insights into emotional clues. Experimental results demonstrate that the extracted spatial and temporal features effectively capture topographic and temporal patterns from cardiac electrical signals, and achieve excellent performance in classification tasks simultaneously. Our finding reveals affective patterns based on body surface potentials for the first time, offering novel insights into affective peripheral patterns analysis.

Upper extremity (UE) impairment is common after stroke resulting in reduced arm use in daily life. A few studies have examined the use of wearable feedback of the quantity of arm movement to promote recovery, but with limited success. We posit that it may be more effective to encourage an increase in beneficial patterns of movement practice-i.e. the overall quality of the movement experience-rather than simply the overall amount of movement. As a first step toward this goal, here we sought to identify statistical signatures of the distributions of daily arm movements that become more prominent as arm impairment decreases, based on data obtained from a wrist IMU worn by 22 chronic stroke participants during their day. We identified several measures that increased as UE Fugl-Meyer (UEFM) score increased: the fraction of movements achieved at a higher speed, forearm postural diversity (quantified by kurtosis of the tilt-angle), and forearm postural complexity (quantified by sample entropy of tilt angle). Dividing participants into severe, moderate, and mild impairment groups, we found that forearm postural diversity and complexity were best able to distinguish the groups (Cohen's D = 1.1, and 0.99, respectively) and were also the best subset of predictors for UEFM score. Based on these findings coupled with theories of motor learning that emphasize the important of variety and challenge in practice, we posit that encouraging people to achieve more forearm postural diversity and complexity might improve the quality of their movement experience and therefore might be therapeutically beneficial.

Sound synthesizers are ubiquitous in modern music production but manipulating their presets, i.e. the sets of synthesis parameters, demands expert skills. This study presents a novel variational auto-encoder model tailored for synthesizer preset interpolation, which enables the intuitive generation of new sounds from pre-existing ones. Leveraging multi-head self-attention networks, the model efficiently learns latent representations of presets, aligning these with perceived timbre dimensions through attribute-based regularization. It is able to smoothly transition between diverse presets, surpassing traditional linear parametric interpolation methods. Furthermore, we introduce an objective and reproducible evaluation method, based on smoothness and non-linearity metrics computed on a broad set of audio features. The model's efficacy is demonstrated through subjective experiments, whose results also highlight significant correlations with the proposed objective metrics. The model is validated using a widespread frequency modulation synthesizer with a large set of interdependent parameters. It can be adapted to various commercial synthesizers, and can perform other tasks such as modulations and extrapolations.

In this publication, we analyze how the performance of propagation parameter estimation for THz channel sounding can be improved by the power spectrum design of a multicarrier waveform. To this end, we discuss the Fisher information of the propagation parameters and the corresponding deterministic Cramèr-Rao lower bound (CRB) as well as their relation to the carrier powers of the excitation signal. We use these quantities to design waveforms that improve range estimation. In practice, optimizing the power spectrum requires prior knowledge of the propagation scenario which is usually not available. Hence, we propose two solutions which we compare numerically to the classical approach of equal power distribution. The numerical evaluation shows that an optimized power distribution can improve the CRB comparable up to a 4 dB gain in SNR depending on whether knowledge about the scenario can be acquired in advance with an additional measurement. © 2023 IEEE.  Personal use of this material is permitted.  Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Status: Accepted for presentation at the EuCAP conference 2024 https://www.eucap2024.org/.

Harmonic retrieval techniques are the foundation of radio channel sounding, estimation and modeling. This paper introduces a Deep Learning approach for joint delay-and Doppler estimation from frequency and time samples of a radio channel transfer function. Our work estimates the two-dimensional parameters from a signal containing an unknown number of paths. Compared to existing deep learning-based methods, the signal parameters are not estimated via classification but in a quasi-grid-free manner. This alleviates the bias, spectral leakage, and ghost targets that grid-based approaches produce. The proposed architecture also reliably estimates the number of paths in the measurement. Hence, it jointly solves the model order selection and parameter estimation task. Additionally, we propose a multi-channel windowing of the data to increase the estimator's robustness. We also compare the performance to other harmonic retrieval methods and integrate it into an existing maximum likelihood estimator for efficient initialization of a gradient-based iteration.

Rapid advancements have been made in artificial intelligence applications recently, and generative models have prominently emerged as effective tools for domain transfer, image enhancement, and simulation. However, when dealing with large-scale gigapixel images, the use of traditional patch-based image aggregation methods introduces checkerboard or blocking artifacts, which compromises image quality and fidelity. Here we propose a parametric kernel that is specifically designed to target the underlying grid structure to mitigate these artifacts. The proposed parametric kernels are validated using three medical imaging modalities for three different generative model tasks, demonstrating improved visual fidelity and quantitative quality evaluation of the generated patch-aggregated images. The proposed method is versatile and compatible with various generative models, offering a robust framework for artifact reduction that can be seamlessly adjusted by modifying kernel parameters, and they can be directly applied and extended to other imaging modalities that employ large-scale images, such as astronomy and satellite imaging. The findings of this study have significant implications for medical imaging applications: by mitigating aggregation artifacts, our approach enhances the overall quality of medical images synthesized with generative models, which is crucial for accurate clinical assessment and subsequent image analysis. Furthermore, the proposed kernels provide a general formulation that can be extended to unpaired tasks, semantic segmentation, classification networks, and other large field-of-view imaging applications.

Full-duplex Cognitive Radio (FD-CR) technology has the potential to significantly improve the spectral efficiency of next-generation wireless systems. However, residual self-interference (RSI), unavoidable in FD systems, represents a colored noise, particularly at mm-wave frequencies. This fact affects blind Spectrum Sensing (SS) algorithms, preventing them from maintaining a constant false alarm rate (CFAR). This causes a power-throughput trade-off resulting in significant performance degradation. Detection in colored noise has been addressed in the literature through time-domain whitening aided by offline training. However, this method is ineffective in the presence of time-varying self-interference (SI). Our paper proposes an adaptive filter-based whitening approach to allow a blind SS to retain the CFAR property in mobile scenarios without the need for offline training. Focusing on low-complexity adaptive filtering, we analytically demonstrate that both the Least Mean Squares (LMS) and the Recursive Least Squares Lattice (RLSL) filters enable the sphericity test to achieve the CFAR property in typical FD-CR scenarios. We highlight the advantages of RLSL over LMS, including faster convergence, superior tracking, and modularity, making it suitable for effective implementation, as well as for efficient low-complexity SI cancellation. Numerical results confirm superior performance in high RSI power scenarios compared to offline solutions and LMS.

In a world where traditional authentication systems are constrained, the introduction of voice-based authentication provides a viable option. This cutting-edge biometric security method uses unique vocal traits including pitch, tone, and speech patterns to create a unique voiceprint. Its sophisticated and dependable verification procedure, when combined with voice-activated services and secure access systems, not only solves au-thentication issues but also improves user experience overall. Our suggested method provides a thorough blueprint for developing a lightweight voice authentication system that is based on text. This system leverages an effective encoder model to convert voice spectrograms, making deployment easy even on edge devices with limited resources. Various dimensionality reduction techniques are explored to obtain optimal voice embeddings that capture speaker uniqueness while minimizing model complexity. A key novelty is the application of model compression techniques including lightweight architectures and Siamese Networks to obtain highly condensed voice embeddings for each user, reducing storage and infrastructure costs compared to traditional voice authentication methods. The proposed lightweight spectrogram embeddings leverage language-agnostic acoustic features, enabling language-independent speaker verification. Additionally, dimensionality reduction applied during voice registration allows the capturing of discriminative voice characteristics in a low-dimensional compact feature space. This significantly cuts down the storage requirements and infrastructure costs per user compared to standard voice biometric approaches, while retaining competitive verification performance. The architecture is optimized for responsiveness by leveraging lightweight frameworks. The proposed system delivers competitive voice authentication capabilities while minimizing memory, computational, and energy footprints. This makes the system useful for integration into smart devices and paves the way for ubiquitous voice biometrics.

The bit error rate (BER) analysis of non-orthogonal multiple access (NOMA) has been widely considered in the literature with the assumptions of perfect and imperfect successive interference cancellation (SIC). For both cases, exact closed-form formulas were derived under various channel models, number of users, and modulation orders. However, all the analysis reported overlooked the transformations that affect the probability density function (PDF) of additive white Gaussian noise (AWGN) after the SIC process. Therefore, the signal model after the SIC process is generally inaccurate, which makes the analysis just approximations rather than exact because the noise after SIC is not Gaussian anymore. Therefore, this letter derives the exact noise PDF after the SIC process and evaluates its impact on the BER analysis. The analytical results obtained show that the noise PDF after SIC should be modeled as a truncated Gaussian mixture. Moreover, the PDF after successful and unsuccessful SIC should be modeled differently.

The rate of car accidents has been increasing in recent years, resulting in losses in human lives, properties and other financial costs. To address this important issue, an embedded machine learning based system is developed. The system is capable of monitoring road conditions, detecting driving patterns, and identifying aggressive driving behaviors. The system is based on neural networks that are trained on a comprehensive dataset of driving events, driving styles, and road conditions. The system can effectively detect potential risks and help mitigate the frequency and impact of accidents. The primary goal is to ensure the safety of both drivers and vehicles. The process of collecting data involved gathering information on three key road events: normal street and normal drive, speed bumps, circular yellow speed bumps, and three aggressive driving actions, sudden starting, sudden stop, and sudden entry. The gathered data is then processed and analyzed using a machine learning system designed for limited power and memory devices. The developed system resulted in a 91.9% accuracy, 93.6% precision and 92% recall.

In recent years, neural network-based black-box modeling of nonlinear audio effects has improved considerably. Present convolutional and recurrent models can model audio effects with long-term dynamics, but the models require many parameters, thus increasing the processing time. In this paper, we propose KLANN, a Koopman-Linearised Audio Neural Network structure that lifts a one-dimensional signal (mono audio) into a high-dimensional approximately linear state-space representation with nonlinear mapping, and then uses differentiable biquad filters to predict linearly within the lifted state-space. Results show that the proposed models match the high performance of the state-of-the-art neural models while having a more compact architecture, reducing the number of parameters by tenfold, and having interpretable components.