Felipe Augusto Pereira de Figueiredo

Classroom behavior analysis has become a relevant area for developing intelligent educational environments, enabling continuous, non-intrusive understanding of student engagement and interactions. In this work, we present a new video clip dataset for recognizing student behavior in the classroom, based on behavior derived from observational protocols. This new dataset also has the feature of obtaining student actions through four different cameras that are positioned frontally and laterally in the classroom. The dataset is then used to train state-of-the-art models for video behavior classification, including Multiscale Vision Transformer (MViTv2-S), Swin Transformer, SlowFast, and X3D. Our results show that MViTv2-S has better performance when accuracy and F1-score are analyzed. The results indicate an accuracy of 74.44 ± 0.45 for the frontal cameras and 70.88% ± 1.04 for the lateral cameras, demonstrating the effectiveness of both camera configurations for student action recognition. When engagement is evaluated by aggregating the classified actions, an accuracy of 83.33% and a macro F1-score of 80.42% are achieved. The results highlight the potential of deep learning-based approaches for classroom behavior analysis, offering insights that may support applications in pedagogical support, student engagement assessment, and the development of interpretable and scalable intelligent educational systems.

Epsilon‐near‐zero (ENZ) media enable unconventional electromagnetic‐wave manipulation through quasi‐uniform field distributions and geometry‐independent responses. Photonic doping extends ENZ functionality by enabling permeability engineering through embedded inclusions while preserving the near‐zero permittivity condition. Although recent developments have expanded photonic doping to resonant metallic, optical, and non‐Hermitian platforms, the perfectly conducting (PEC) limit has remained largely unexplored for radiating ENZ structures. Here, the PEC limit is systematically developed as a passive, nonresonant approach for permeability engineering in waveguide‐based ENZ cavities with radiating apertures. Analytical derivations and full‐wave simulations show that a solid metallic inclusion suppresses the magnetic field within its volume, yielding an effective permeability governed by the ratio between the excluded and cavity cross‐sectional areas. The resulting response is invariant with respect to inclusion geometry, position, and multiplicity, provided the total excluded area remains unchanged. The proposed implementation reproduces the impedance‐matching and radiation characteristics of dielectric and resonant photonic dopants while achieving higher radiation efficiency without resonant energy storage or engineered capacitive loading. These results demonstrate that dielectric, resonant metallic, and PEC inclusions represent different implementations of the same photonic‐doping framework, establishing the PEC limit as a practical low‐loss methodology for designing efficient ENZ cavities and antennas.

Background: Peripheral nerve sheath tumors (PNSTs) of the head and neck (H&N) show histopathological overlap. Although convolutional neural networks (CNNs) have demonstrated feasibility in soft tissue tumor classification, limited intra-class variability related to perineurioma remains a critical constraint for rare tumor subtypes. Methods: This retrospective diagnostic accuracy study with internal validation included 30 patients diagnosed with PNSTs. Whole-slide images were digitized at 20× magnification and partitioned using a strict patient-wise split. Synthetic perineurioma patches were generated using a modified Pix2Pix-based Generative Adversarial Network (GAN) incorporating a bottleneck architecture and self-attention modules. Two morphology-driven augmentation strategies were evaluated: (1) intra-phenotypic expansion by cross-patient patch pairing within the sclerosing subtype and (2) inter-phenotypic interpolation by cross-phenotype patch pairing between sclerosing and intraneural variants. EfficientNetV2-B0 pre-trained on ImageNet was trained under three configurations: original dataset only, original + Experiment A synthetic patches, and original + Experiment B synthetic patches. All performance metrics were computed exclusively on an independent yet internal test set composed of original images. Results: GAN-based augmentation improved global classification performance compared with the baseline model trained on original images only (accuracy 0.733). Intra-phenotypic expansion increased accuracy to 0.767 and achieved the highest balanced accuracy (0.750) and macro-F1 (0.740). Inter-phenotypic interpolation yielded the highest overall accuracy and competitive multiclass agreement metrics. Perineurioma recall improved from 0.34 (baseline) to 0.51 with intra-phenotypic augmentation and 0.47 with inter-phenotypic interpolation, while specificity remained ≥ 0.999 across all strategies. Conclusions: Structured, pathology-informed GAN augmentation improved CNN classification of PNSTs, particularly for the morphologically heterogeneous perineurioma class. Intra-phenotypic expansion primarily improved rare-class sensitivity, whereas inter-phenotypic interpolation improved multiclass agreement and global robustness. These findings support morphology-driven synthetic enrichment as a clinically meaningful strategy to improve AI performance in underrepresented tumor entities and potentially support diagnostic decision-making in digital pathology environments.

This paper presents a novel framework for analog variance noise modulation (VNM) and its corresponding variance-based demodulation technique. Unlike traditional digital noise communication systems such as Kirchhoff–Law–Johnson–Noise (KLJN) schemes that encode binary data using discrete noise levels, the proposed analog VNM continuously maps a signal’s amplitude onto the variance of a white Gaussian noise (WGN) carrier. In this way, information is embedded within the statistical behavior of noise rather than its waveform, enabling inherently secure and noise-like transmission. To accurately reconstruct the transmitted signal, we develop a new state-space modeling and inference framework inspired by the Discrete Fourier Transform (DFT). This model captures the temporal structure of the analog baseband signal through its harmonic components while treating the noise variance as a time-varying latent process. Building upon this structure, a Rao–Blackwellized Particle Filter (RBPF) is designed to estimate the underlying signal by jointly inferring its harmonic dynamics and the hidden variance states, enabling robust demodulation under stochastic noise conditions. Simulation results demonstrate that the proposed analog VNM achieves accurate signal recovery and exhibits strong concealment properties compared to classical amplitude and frequency modulation schemes. Overall, this work introduces a new class of analog noise-based communication with an inherently simple and energy-efficient transmitter, while employing a higher-complexity receiver-side inference framework for accurate signal recovery. This asymmetric architecture makes the proposed VNM particularly suitable for low-power sensing and Internet of Things (IoT) scenarios where ultra-lightweight transmitters communicate with computationally capable gateways or base stations.

Selecting the Modulation and Coding Scheme (MCS) from the measured Signal-to-Noise Ratio (SNR) is commonly implemented via precomputed Look-Up Tables (LUTs) derived from analytical models, link-level simulations under idealized assumptions, or empirical Bit Error Rate (BER)/Block Error Rate (BLER) thresholds. In operational Software-Defined Radio (SDR) platforms, however, non-ideal channel estimation, pilot-grid interpolation, equalizer noise enhancement, decoder non-idealities, and Radio Frequency (RF) front-end impairments distort this mapping, causing conventional tables to miss the expected goodput. This paper proposes a measurement-driven Adaptive Modulation and Coding (AMC) strategy based on Contextual Bandit (CB), which learns a stationary SNR→MCS policy directly from link measurements by maximizing downlink Medium Access Control (MAC)-layer goodput. The target application is a TV White Space (TVWS) backhaul in the Ultra High Frequency (UHF) band, designed to provide Internet connectivity to remote areas: both the Base Station (BS) and the remote terminal are fixed, typically under Line-of-Sight (LoS), yielding quasi-stationary channel conditions over many transmission intervals. We implement and evaluate the proposed approach on an SDR-based radio access system in two representative deployments: (i) an RF wireless link that exercises the full Physical Layer (PHY) stack under regulatory transmit-power constraints, and (ii) a Radio over Fiber (RoF) extension enabling tens-of-kilometers reach. The context is the measured SNR (discretized into bins), the action is the MCS index, and the reward is the delivered goodput; transmission errors and retransmissions are naturally captured through reduced goodput. For controlled experimentation and reproducible operating regimes, we constrain the effective SNR within a desired interval by programmatically adjusting the transmit-chain gain. As baselines, we consider as references the link adaptation based on an Inner Loop Link Adaptation (ILLA) obtained from laboratory calibration, an Outer Loop Link Adaptation (OLLA) that periodically adjusts the selected MCS using a BER target, and a supervised Histogram-Based Gradient Boosting Regressor (HGBR) baseline, trained offline from measured samples to approximate the goodput-maximizing SNR→MCS mapping. Experimental results show that the learned CB policy improves the area under the goodput–SNR curve and the mean Spectral Efficiency (SE) over both the baselines within the operational SNR range, while satisfying the BER target. In addition, by directly learning the optimal SNR→MCS mapping from interaction with the real environment, the proposed approach avoids the approximation stage required by the offline HGBR model and eliminates the transient convergence period required by the BER-driven OLLA to adapt from the initial ILLA table toward its steady-state operating point, thereby enabling immediate near-optimal performance in quasi-stationary wireless backhaul links.

Cognitive radio systems operating in dynamic environments face severe performance degradation due to composite fading and user mobility. While feature-based sensing techniques, such as cosine feature recognition detection (CRD), offer improved robustness, the existing literature lacks a unified analytical framework to characterize its performance under generalized fading and mobility conditions. This paper investigates the robustness of CRD under generalized fading and mobile environments. First, we derive a novel closed-form expression for the average probability of detection using the bivariate Fox H-function. This formulation provides a unified solution for diverse fading scenarios, which we apply to a composite channel model incorporating α-μ small-scale fading, Gamma-distributed shadowing, and random waypoint mobility. Second, to facilitate practical system design, we propose an efficient algorithm to estimate the minimum number of sensing samples required to meet stringent quality-of-service requirements in modern wireless networks. Third, to address the limitations of black-box machine learning (ML), we investigate the feasibility of a novel symbolic detector and compare it with some ML-based detectors. Unlike computationally intensive neural networks, this data-driven approach discovers a compact, interpretable mathematical expression that combines CRD and energy detection (ED) statistics. Simulations validate the theoretical analysis and demonstrate that the proposed symbolic detector, along with the ML-based ones, outperform traditional ED and CRD, particularly at low signal-to-noise ratios and in harsh mobile scenarios.

The sum of Gamma-Gamma (also known as Generalized-K) random variables (RVs) is fundamental in wireless and radar applications, yet its statistical characterization remains challenging in terms of analytical tractability, implementability, and accuracy. This letter derives exact, lightweight expressions for the probability density function (PDF) and cumulative distribution function (CDF) of the sum of independent and identically distributed Gamma-Gamma RVs. Unlike the prior— and only—exact series expansion for the sum PDF proposed by Bayaki et al. [1], which entails a combinatorial complexity of O((L+1)(^I+L_L))), the series representation developed herein reduces the computational burden to O(LI³), where L denotes the number of Gamma-Gamma RVs in the sum and I the series truncation order. Numerical results further demonstrate substantial runtime savings over [1]. For example, for L = 30, our method achieves more than a 100× speedup. For benchmarking, widely used approximations (e.g., Gamma, Gamma-Gamma, and α-μ) are also considered, but they fail to accurately capture the distribution tails, particularly for large values of L.

Fluid antenna systems represent an innovative approach in wireless communication, recently applied in multiple access to optimize the signal-to-interference-plus-noise ratio through port selection. This paper frames the port selection problem as a multi-label classification task for the first time, improving best-port selection with limited port observations.We address this challenge by leveraging liquid neural networks (LNNs) to predict the optimal port under emerging fluid antenna multiple access scenarios alongside a more general 𝛼-𝜇 fading model. We also apply hyperparameter optimization to refine LNN architectures for different observation scenarios. Our approach achieves outage probability performance among the best-performing methods in the literature.