Optical polarizers, which selectively transmit light with specific polarization states, are essential components in modern optical systems. Here, we experimentally demonstrate integrated waveguide and microring resonator (MRR) polarizers incorporating reduced graphene oxide (rGO). 2D graphene oxide (GO) films are integrated onto silicon photonic devices with precise control over their thicknesses and sizes, followed by GO reduction via two different methods including uniform thermal reduction and localized photothermal reduction. We measure devices with different lengths, thicknesses, and reduction degrees of the GO films. The results show that the devices with rGO exhibit better performance than those with GO, achieving a polarization-dependent loss of ~47 dB and a polarization extinction ratio of ~16 dB for the hybrid waveguides and MRRs with rGO, respectively. By fitting the experimental results with theory, it is found that rGO exhibits more significant anisotropy in loss, with an anisotropy ratio over 4 times that of GO. In addition, rGO shows higher thermal stability and greater robustness to photothermal reduction than GO. These results highlight the strong potential of rGO films for implementing high-performance polarization selective devices in integrated photonic platforms.

In this paper, an orthogonal mode decomposition method is proposed to decompose finite length real signals on both the real and imaginary axes of the complex plane. The interpolation function space of finite length discrete signal is constructed, and the relationship between the dimensionality of the interpolation function space and its subspaces and the band width of the interpolation function is analyzed. It is proved that the intrinsic mode is actually the narrow band signal whose intrinsic instantaneous frequency is always positive (or always negative). Thus, the eigenmode decomposition problem is transformed into the orthogonal projection problem of interpolation function space to its low frequency subspace or narrow band subspace. Different from the existing mode decomposition methods, the orthogonal modal decomposition is a local time-frequency domain algorithm. Each operation extracts a specific mode. The global decomposition results obtained under the precise definition of eigenmodes have uniqueness and orthogonality. The computational complexity of the orthogonal mode decomposition method is also much smaller than that of the existing mode decomposition methods.

Synthetic Aperture Radar (SAR) systems are increasingly vulnerable to Radio Frequency Interference (RFI), creating a demand for advanced detection and mitigation strategies in complex operational environments. This paper introduces a novel framework for segmenting portions of the raw SAR data affected by RFI. Given the absence of publicly available SAR datasets with labels, spaceborne SAR data is simulated using SEMUS, an open-source RF-level SAR simulator, emulating Sentinel-1 raw data. This serves as the training data for a Deep Convolutional Neural Network (DCNN) optimized to segment the RFI portions. Performance is quantitatively validated against the classical energy detection method for interference detection. The robustness of the framework is further qualitatively tested on real Sentinel-1 SAR data in a zero-shot manner, demonstrating its capability to generalize to real-world scenarios despite the absence of real labeled data. This framework offers an adaptable solution for RFI detection in spaceborne SAR applications, contributing to improved data quality for Earth observation.

In this article, we propose a digital agent (DA)assisted network management framework for future sixth generation (6G) networks considering users' quality of experience (QoE). Particularly, a novel QoE metric is defined by incorporating the impact of user behavior dynamics and environment complexity on quality of service (QoS). A two-level DA architecture is developed to assist the QoE-driven network orchestration and slicing, respectively. To further improve the performance of proposed framework, three potential solutions are presented from the perspectives of DA data collection, network scheduling algorithm selection, and DA deployment. A case study demonstrates that the proposed framework can effectively improve users' QoE compared with benchmark schemes.

This paper presents a method for localization of a time-harmonic source in a cavity with magnitude-only measurements using the principles of electromagnetic time reversal. A time-harmonic point source that radiates an electromagnetic field at a specific frequency is considered. The source is assumed to be at an unknown position within the cavity. Scatterers of known geometry and dielectric constants present in the cavity alter the source field. The magnitude of the electric field is measured at certain locations within the cavity. Using these magnitude measurements, the position of the source is determined. Linear constraints based on electromagnetic time reversal are derived. These constraints depend only on the magnitude of electric fields. A convex optimization problem minimising the L1 norm subject to the derived linear constraints is formulated. The Matching pursuit algorithm is used to solve the optimization problem. The solution to this optimization problem reveals the source location. The proposed method is validated by numerical experiments in both 2D and 3D cavities. Numerical results suggest that it is possible to localize a time-harmonic point source with three magnitude-only measurements (despite having many scatterers).

Evaluating privacy in data-driven software is hindered by its subjective nature and the multitude of privacyenhancing technologies. To address this, we propose a framework for categorizing data based on its privacy implications. This proposal aims to establish a standardized approach for quantifying privacy across various data sources, aligned with DevPrivOps.

Physical layer security is a promising technology for non-terrestrial wireless networks due to the lack of infrastructure over the low-altitude and high-altitude uncrewed aerial vehicle platforms. Physical layer security can be addressed using key-based and keyless approaches, such as channel adaptation and channel coding. In the key-based approach, the legitimate users generate secret keys by exploiting the unique random channel gains, which can be further used for information encryption and decryption. In non-terrestrial networks, the physical layer exploitation and the matching of generated secret keys on legitimate ends become more complex due to the 3-dimensional channel models, and the high mobility of uncrewed aerial vehicles, resulting in nonzero key. To address this challenge, this paper proposes an artificial intelligence-based key exchange mechanism where the secret keys are generated on the legitimate nodes of the communications links. The transmitter is responsible for training the neural network and transmitting the final synapses to the receiver to reconcile the information. The results are presented in terms of secret key waveforms. The success rate against the number of errors that occur in a single 128-bit key and 256-bit key shows a significant improvement.

Bionic limb control through myoelectric pattern recognition, offering intuitive decoding of motor intent, can improve the quality of life for individuals with amputations. However, most work on pattern recognition only uses a small subset of the myoelectric data generated during daily life, to train an Artificial Neural Network (ANN) via Supervised Learning (SL). Scenarios substantially different from the recording session e.g. different limb positions, can lead to misclassifications by the ANN during everyday usage of the bionic limb. Recording labeled data from all scenarios encountered in daily life could alleviate the problem, but would be prohibitively time consuming. Unsupervised Domain Adaptation (UDA) offers a solution by leveraging unlabeled data from a target domain, not represented in the labeled dataset i.e. the source domain, to calibrate ANNs for improved performance. In this study we explore the potential of two UDA algorithms for domain shifts in myoelectric pattern recognition: Domain Adversarial Neural Networks (DANN) and Sliced Wasserstein Discrepancy (SWD). Offline evaluation identified SWD as the best-performing algorithm, which was subsequently validated in online experiments with 11 participants. Using UDA substantially improved the performance, not only on the target domain, but also on the source domain. Indeed, it nearly matched the performance of an ANN trained through SL on labeled data from both the source and target domain. This is the first time UDA was validated in online experiments as a viable approach to overcome domain shift problems caused by changes in limb position.

Imaging spectroscopy and lightweight unmanned aerial vehicles (UAVs) have revolutionized remote sensing of vegetation, by providing high spectral and spatial resolution data. Comprehensive wetland monitoring should be based on reliable mapping of biophysical aquatic vegetation parameters, which in turn requires low bias in measured spectra, e.g. minimization of reflectance anisotropy. This study aims to quantitatively investigate how sensor scan direction and solar zenith angle (SZA) affect vegetation spectra over two dominant aquatic plants, Phragmites australis and Nuphar lutea, representing different functional types (i.e. emergent and floating) with divergent canopy structure and affinity with water, using data collected with a hyperspectral (400-1000 nm range) push-broom sensor mounted on a UAV. Anisotropy factor (ANIF), mean reflectance scatterplots and regression analysis were used for assessing the magnitude of spectral anisotropy at both pseudo-leaf and canopy scales. Our findings showed more accentuated anisotropic behaviour in the visible than near-infrared domain at SZA<60° as the scan direction approaches that of solar principal plane. The increase in reflectance in near-forward direction affects the spectra of floating-leaved N. lutea at both leaf and canopy scales, suggesting the presence of water (as a film over leaves or as canopy background) as a likely anisotropy driver. Backward hotspot is evident in canopy spectra of emergent P. australis, with driving mechanisms similar to terrestrial species (i.e. shadow-hiding). At SZA>60°, reflectance anisotropy appears to be negligible, independent of functional type and scan direction. This research underscores the importance of considering sun-target-sensor geometry in remote sensing studies on aquatic vegetation.

The trend towards sensor miniaturization has heightened interest in optimizing sensor array configurations across various scientific and industrial applications that imply multichannel measurements. In this article, we present an extended version of our previous approach, adapting it to the realworld conditions in magnetoencephalography (MEG). At each step the method selects the sensor location that maximizes the ROI-related signal-to-noise ratio (SNR), followed by the projection recursively applied to the gain matrix to orthogonalize the subsequent stereotypic steps with respect to the source subspace served by the already selected sensors. This algorithm offers high computational efficiency and flexibility with respect to physical constraints. We dub our approach as Optimal Layout Design via Recursively Applied Leadfield Elimination (OLD RALFE). Our analysis of the optimized MEG sensor layouts obtained with OLD RALFE shows that the proposed algorithm performs similarly to several other significantly more computationally demanding sensor layout optimization approaches in terms of both SNR and the Cramér-Rao lower bound (CRLB) on the spatial resolution of the inverse problem solution. OLD RALFE demonstrated in application to MEG, is broadly applicable to many other multichannel measurement challenges.

Hardware accelerators for artificial intelligence (AI) applications must often meet stringent constraints for accuracy and throughput. In addition to architecture/algorithm improvements, high performance computational techniques such as mixed precision are also required. In this paper, a floating-point fused multiply-add (FMA) unit supporting mixed/multiple precision is proposed. A wide range of conventional precision formats (such as half and single, formats) as well as emerging precision formats (including E4M3, E5M2, DLFloat, BFLoat16 and TF32) are supported in the proposed design. In addition to all these formats, the proposed design is flexible in manipulating the exponent and mantissa lengths for 8, 16 and 32-bit floating-point numbers based on the needs of an application. The proposed FMA can be configured to perform normal FMA operation supporting multiple precision, or alternatively perform mixed precision in ASIC. The proposed FMA is fully pipelined and in each cycle, the input bit streams are processed based on the provided configuration, so independent of the previous cycles. For normal FMA, the proposed design utilizes sharing of resources to parallelize several operations based on the available hardware and required precision. Mixed precision tends to accumulate the lower precision dot products into higher precision to avoid overflow/underflow. The proposed design improves computational accuracy by adding all possible dot products at the same time while decreasing the number of rounding operations to prevent rounding errors. An innovative method to accumulate the dot products and the aligned addend is also proposed for normal/mixed precision FMA operation. By considering tradeoffs between reusing the available hardware and removing unnecessary complex units, a more efficient and flexible design is attained in terms of hardware metrics and supported mixed precision computation compared to other designs found in the technical literature. Extensive simulation results for a comparative analysis are provided.

Adders are the most important element in the ALU. Speeding the addition improves the performance of the ALU in particular and the CPU in general. Pipelining is used to process tasks in parallel with minimal adjustment of the hardware needed to process the same tasks sequentially. In this paper, a pipelining procedure is presented to enable the half adder to perform the addition of two numbers in parallel. The resulted speed up factor is guaranteed to perform the same addition in half the clock cycle of full adder.

A general method to construct wavelet function on real number field is proposed in this article,which is based on finite length sequence.This finite length sequence is called the seed sequence, and the corresponding wavelet function is called the seed sequence wavelet function.The seed sequence wavelet function is continuous and energy concentrated in both time and frequency domains.That is, it has a finite support set in both time and frequency domains. It is proved that if and only if the seed sequence has 0 mean value, the interpolation function satisfies the admissible condition of wavelet function.The conditions corresponding to the higher order vanishing moment of the seed sequence wavelet function are also given in this article. On this basis, the concept of random wavelet function is proposed, and the condition of the regularity of random wavelet is discussed.

There are infinitely many primes which are of the form in near-square.

Although deep learning (DL) has led to several breakthroughs in many disciplines as diverse as chemistry, computer science, electrical engineering, mathematics, medicine, neuroscience, and physics, a comprehensive understanding of why and how it is empirically successful remains fundamentally elusive. To attack this fundamental problem and unravel the mysteries behind DL's empirical successes, the DL research community has made significant innovations in optimization, generalization, and approximation. However, no work to date has quantified the testing performance of a DL-based algorithm employed to solve a pattern classification problem. To address this fundamental challenge in part, we expose the testing performance limits of binary classifiers based on deep rectified linear unit (ReLU) feedforward neural networks (FNNs) that are trained with hinge loss by driving their fundamental asymptotic testing performance limits. These limits are validated by our extensive computer experiments.

This study proposes an innovative hierarchical binary classification approach based on scalp EEG data to predict comatose patients' Glasgow Outcome Scale (GOS).The dataset consists of continuous EEG recordings from 13 comatose patients classified into three GOS groups: GOS 1 (death), GOS 3 (severe disability), and GOS 5 (good recovery). Traditional machine learning (ML) and deep learning (DL) models, though effective in some instances, struggle to generalize across diverse patient populations, mainly when evaluated with leave-one-subject-out (LOSO) cross-validation. We introduce a two-step binary classification approach to address these challenges, utilizing Convolutional Neural Networks (CNNs) with spectrogram features for one classifier and Random Forest (RF) with Approximate Entropy (ApEn) for the other. By simplifying the multi-class classification task into two binary decisions, we achieve perfect classification performance for GOS 1 and GOS 5, while GOS 3 is inferred by exclusion. Importantly, we found that just 2 minutes of EEG data per subject was sufficient to achieve this level of performance, offering significant implications for clinical applications where data is often limited. Even considering the small dataset, our results prove that our method performs better than traditional approaches, providing a scalable, efficient solution for coma outcome prediction.

The rapid evolution of language models has significantly enhanced their ability to generate human-like text; however, challenges persist in achieving complex syntactic and contextual understanding. Introducing Dynamic Syntax Transfer (DST), a novel methodology that dynamically adjusts syntactic structures in response to contextual variations, thereby enabling large language models to produce text with improved grammatical accuracy and contextual relevance. The implementation of DST within an open-source language model architecture has demonstrated notable advancements in syntactic precision and contextual coherence, as evidenced by comprehensive quantitative and qualitative analyses. These findings demonstrate DST's potential to advance context-aware syntax modeling, offering a promising avenue for future research and applications in sophisticated language generation systems.

Generating executable code from natural language instructions using Large Language Models (LLMs) poses challenges such as semantic ambiguity and understanding task-specific contexts. To address these issues, we propose a system called _DemoCraft_, which enhances code generation by leveraging in-context learning and demonstration selection, combined with latent concept learning. Latent concept learning introduces additional concept tokens, which are trainable embeddings that capture task-specific knowledge. We then test our system on two major datasets: _MBPP_ and _Humaneval_. Our experimental results demonstrate that the proposed system achieves an approximate _2x_ increase in the _pass@k_ metric compared to baseline models. Furthermore, we introduce two novel evaluation metrics: _correctness@k_ and _similarity@k_. Our empirical studies indicate that our system attains nearly a _3x_ improvement in these metrics as well.