בית הספר
להנדסת חשמל ומחשבים

Abstract: Understanding how large-scale brain networks represent visual categories is fundamental to linking perception and cortical organization. Using high-resolution 7T fMRI from the Natural Scenes Dataset, we construct parcel-level functional graphs and train a signed Graph Neural Network that models both positive and negative interactions, with a sparse edge mask and class-specific saliency. The model accurately decodes category-specific functional connectivity states sports, food,vehicles) and reveals reproducible, biologically meaningful subnetworks along the ventral and dorsal visual pathways. This framework bridges machine learning and neuroscience by extending voxel-level category selectivity to a connectivity-based representation of visual processing. Bio: Shira Karmi is Electrical Engineer MSc student supervised by Tammy Riklin Raviv and Galia Avidan Abstract: Accurate delineation of individual cells in microscopy videos is essential for studying cellular dynamics, yet separating touching or overlapping instances remains a persistent challenge. Although foundation-model for segmentation such as SAM have broadened the accessibility of image segmentation, they still struggle to separate nearby cell instances in dense microscopy scenes without extensive prompting. We propose a prompt-free, boundary-aware instance segmentation framework that predicts signed distance functions (SDFs) instead of binary masks, enabling smooth and geometry-consistent modeling of cell contours. A learned sigmoid mapping converts SDFs into probability maps, yielding sharp boundary localization and robust separation of adjacent instances. Training is guided by a unified Modified Hausdorff Distance (MHD) loss that integrates region- and boundary-based terms. Evaluations on both public and private high-throughput microscopy datasets demonstrate improved boundary accuracy and instance-level performance compared to recent SAM-based and foundation-model approaches.

Sensing of biological interactions with transistor-based biosensor Point-of-care (POC) diagnostic technologies play a critical role in enabling rapid clinical decision-making and improving patient outcomes. Despite this potential, their widespread implementation remains limited by the absence of platforms that simultaneously offer miniaturization, high accuracy, multiplexed detection from ultra-small sample volumes, and cost-effective operation. Transistor-based biosensors present a compelling candidate for POC applications, owing to their inherent scalability, robustness, sensitivity, and ultimate potential for sensing of multiple targets in ultra small samples. However, field-effect transduction of biomolecular events remains an unresolved challenge due to the inherent localization and non-uniform distribution of such nano-scaled biomolecular events, as well as operation in solution environments. We developed the Meta-Nano-Channel biological field-effect transistor (MNC bioFET) addressing the challenges of localized and solution gating. The MNC architecture enables the formation of multiple conducting nano-channels scanning for molecular events while maintaining the solution and molecular species in electrochemical equilibrium. Here, we demonstrate the MNC bioFET for specific, label-free, quantitative, and real-time detection of C-reactive protein (CRP)—a key biomarker of inflammation and tissue damage—directly in 0.5 μL unprocessed blood and sweat samples.

Abstract: Maor Kakun This research investigates emotion recognition using Remote Photoplethysmography (rPPG) signals extracted from facial videos. rPPG estimates the blood volume pulse by tracking subtle temporal variations in facial pixel intensities, enabling non-contact physiological measurement. Extracting high-quality rPPG signals under motion artefacts, video compression, and challenging illumination conditions remains difficult. Current state-of-the-art approaches rely on large neural network models that operate as black boxes, directly mapping video frames to rPPG signals with high computational cost and limited interpretability. This study proposes a hybrid framework that combines classical signal processing techniques, physically interpretable feature extraction, and a lightweight deep learning architecture. The proposed approach reduces computational complexity while improving model explainability and robustness across varying recording conditions. Using the resulting high-quality rPPG signals, reliable heart rate variability features are obtained, enabling effective emotion recognition. Abstract: Fatema Abu Elheija Rapid detection of toxic chemical species is important for environmental monitoring and chemical safety. In this work, we demonstrate label-free, real-time, and quantitative detection of sodium hypochlorite solution (SHC) using a Meta-Nano-Channel Chemical Field-Effect Transistor (MNC ChemFET) in PBS solution. The SiO₂ sensing surface is functionalized with an indole-based recognition layer that selectively reacts with hypochlorite molecules, producing changes in interfacial surface charge that modulate the transistor’s electrical response. The multi-channel MNC structure enhances electrostatic coupling and improves sensitivity and signal stability under electrolyte conditions. Calibration and control measurements confirm quantitative sensing performance, selectivity, and operational stability. Abstract: Shani Budilovsky Automatic Speaker Verification (ASV) is a biometric authentication technology that verifies a speaker’s identity based on their voice characteristics and is widely used in security-sensitive applications such as access control and identity verification. Despite their success, ASV systems are vulnerable to spoofing attacks, including replay, synthetic, and manipulated speech. To address this challenge, Spoofing-robust ASV (SASV) systems combine speaker verification models with dedicated spoofing countermeasures to jointly improve verification accuracy and attack resistance. Data augmentation has become a key technique for improving the generalization and robustness of deep learning-based ASV and countermeasure models by increasing training data diversity and simulating real-world acoustic variability. Augmentation methods can be applied at the waveform level or at the time-frequency representation level, such as spectrogram-based techniques. However, in the SASV setting, not all augmentation strategies are equally beneficial, and some may improve speaker discrimination while degrading spoof detection performance, or vice versa. This seminar examines the effect of different data augmentation strategies on both ASV and spoofing-robust ASV systems, and discusses the importance of carefully designing augmentation pipelines that balance speaker verification accuracy with spoofing robustness in practical SASV systems

The maximum continuous-wave (CW) output power achievable from fibre lasers typically decreases exponentially with increasing wavelength up to approximately 3 µm, primarily due to the quantum defect, defined by the ratio of laser photon energy to pump photon energy. However, until about a decade ago, fibre lasers operating beyond 3 µm were limited to output powers of only ten milliwatts well below the trend dictated by their wavelengths. The introduction of the dual-wavelength pumping (DWP) technique by researchers at the University of Adelaide enabled efficient laser emission around 3.5 µm from the 4F9/2 → 4I9/2 transition in erbium-doped fluoride fibres, such as ZBLAN. Using this approach, our team demonstrated more than an order-of-magnitude increase in CW output power and efficiency, followed by a further order-of-magnitude improvement to Watt levels, that not only reached but surpassed the long-established exponential power-scaling trend. The success of DWP has revitalised the previously niche field of mid-infrared fibre lasers, leading to the emergence of numerous research groups pursuing fibre laser development at wavelengths beyond 3 µm and the opening of commercial applications. This presentation is divided into three parts. First, I will introduce the fundamentals of the DWP technique and trace its development. I will then present recent advances in power scaling and short-pulse operation. Finally, I will discuss our ongoing work on the use of 3.5 µm fibre lasers for the thermal compensation system of LIGO, with the goal of enabling higher-power operation and extending the detection range for gravitational-wave events.

Seminar Abstract Naor Balas Protein–DNA interactions play a pivotal role in essential biological processes such as replication, repair, and recombination. Monitoring these interactions, and their disruption by small molecules, is vital for advancing molecular biology and developing innovative therapeutic strategies. However, conventional techniques often rely on labeling, indirect readouts, or slow and complex protocols, limiting their real-time applicability and sensitivity. This seminar presents a novel approach that employs a Meta-Nano-Channel biological Field-Effect Transistor (MNC bioFET) for real-time, label-free, and highly specific detection of DNA–primase interactions. The CMOS-compatible silicon-on-insulator (SOI) device, functionalized with DNA probes or primase, enables direct electrical readout of binding and dissociation events through modulation of the source–drain current. By introducing indole-derivative molecules, the platform further demonstrates quantitative analysis of small-molecule-induced disruption of primase–DNA binding. Extensive control experiments confirm the specificity and robustness of the method. This is the first reported use of MNC bioFET technology to simultaneously detect protein–DNA interactions and their disruption by small molecules in real time. The findings highlight the potential of this platform as a tool for molecular interaction studies, drug screening, and the development of next-generation label-free biosensing technologies. Keywords: Protein–DNA interactions, Primase, BioFET, Label-free biosensor, Real-time detection, Small molecule disruption Seminar Abstract: Chay Aflalo This seminar presents research on improving anomaly detection in hyperspectral images using a refined implementation of the RX algorithm. Hyperspectral images provide rich spectral information for each pixel, enabling the detection of subtle targets that are not visible in standard imaging. However, the high dimensionality and redundancy of the data pose major challenges for efficient and reliable analysis. The work investigates how different data manipulation and processing strategies such as dividing the hyperspectral cube into sub-cubes, applying noise filtering, and using alternative background estimation methods affect the performance of the RX algorithm. The study includes experiments on real hyperspectral datasets, such as the Viareggio and RIT datasets, and evaluates how segmentation and covariance modeling influence anomaly scores. The goal of the research is to enhance both the accuracy and robustness of anomaly detection, while also reducing computational complexity. The results contribute to improved hyperspectral image processing methods with applications in environmental monitoring, agriculture, and defense. Abstract: Shadad Watad Characterisation of nonlinear optical properties of MXene for photonic application, and implementing all optical activation function of neural networks based on waveguide MXene. Different geomtries of MXenes are investigated to study how the activation function changes.

Abstract Organic semiconductors are central to modern electronic and optoelectronic devices. While thin-film morphology critically governs charge transport and device performance, common deposition methods provide limited morphological control, necessitating sensitive characterization techniques to elucidate the physical properties. We demonstrate that spectroscopic ellipsometry (SE) combined with Critical Point (CP) analysis enables rapid, non-destructive assessment of microstructural evolution during thermal annealing. The CP model provides physically meaningful optical parameters that exhibit strong correlations with electrical performance. Using P3HT-based organic thin-film transistors as a benchmark system and extending the approach to other polymers and device architectures, our results reveal a clear annealing-induced dimensionality shift in the optical transition, indicating increased crystallinity and reduced energetic disorder. These optically derived descriptors accurately predict improvements in charge-carrier mobility, establishing SE–CP analysis as a practical tool for early-stage materials evaluation and process optimization.

Abstract Optimal control is a widely used approach for controller synthesis in numerous disciplines, including epidemiology (intervention policies), biological engineering and synthetic biology (design of bio-molecular circuits), aerospace engineering (trajectory optimization) and robotics and automation (motion planning). In such applications, dynamical models that account for delays in the evolution of the system (i.e. time-delay systems - TDSs) are ubiquitous; for example, an epidemiological model can take into account the incubation period of a disease. Compared to regular ODEs, optimal control of TDSs imposes significant additional challenges in synthesizing controllers and designing reliable algorithms with theoretical guarantees to identify optimal controllers. In this talk, I will first present the general problem of optimal control of TDSs and discuss some of its analogies with the problem of minimizing functions. Then, I will present a recently developed sequential algorithm for optimal control of TDSs, and discuss its guarantees. Finally, I will show numerical simulations that demonstrate the efficacy of the algorithm in solving problems emanating from mathematical epidemiology. The talk will require no prior knowledge of optimal control, and all relevant concepts will be introduced throughout the talk. Bio Rami Katz is an Assistant Professor at the School of Electrical and Computer Engineering at Tel Aviv University. He holds a B.Sc. and M.Sc. in Applied Mathematics and a Ph.D. in Electrical Engineering, with specialization in Systems and Control. His research interests revolve around Systems and Control and Dynamical Systems, and their applications to data analysis, inverse problems and systems biology. Rami is the recipient of multiple awards and prizes, including the Automatica Editor’s choice award in 2020, the Juan De La Cierva fellowship in 2023 and the Alon Fellowship in 2025. He is an associate editor of the European Control Conference and of the journal Mathematics of Control, Signals and Systems.

אוניברסיטת בן-גוריון בנגב גאה להזמין לטקס הענקת תואר ראשון ושני של הפקולטה למדעי ההנדסה

First Speaker: Omer Hazan MSc candidate supervised by Nir Shlezinger Abstract: Radio frequency fingerprinting identification (RFFI) enables device authentication by leveraging unique hardware- induced imperfections in analog signals. While recent deep learning- based methods have shown promise for RFFI, they often target only a subset of tasks such as closed-set classification or open-set recog- nition, and do not support discovering new devices. In this work, we propose a unified and extensible framework for RFFI based on deep learning-aided vector quantization. Our method casts feature extraction as a vector quantization problem in a learned latent space, enabling all three RFFI tasks through a single architecture: classifi- cation is performed via codeword selection, detecting unseen devices is achieved through entropy-based rejection, and registering them is realized by dynamically expanding the codebook with new entries. Evaluations on LoRa signal datasets demonstrate that our method outperforms existing RFFI solutions in accuracy, while uniquely enabling scalable and unsupervised discovery of new devices. Index Terms—RF fingerprint identification, device discovery. Bio: Omer Hazan MSc candidate @ BGU, supervised by Nir Shlezinger Second Speaker: Raz Zohar Electrical engineering MSc candidate supervised by Nir Shlezinger Abstract: Remote localization from array measurements is central to radar, wireless sensing, and autonomous systems, but practical deployments often involve multiple spatially separated sensors that must transmit limited information to a remote processor. In this seminar, we present a task-oriented remote inference framework for direction-of-arrival estimation that combines deep learning with classical array signal processing. Each sensing device extracts compact subspace-oriented features from its local observations, enabling the remote server to fuse information across multiple arrays and recover source directions under communication constraints. The framework preserves the structure of interpretable subspace methods such as ESPRIT and MUSIC while improving robustness to challenging conditions, including coherent sources, low SNR, limited snapshots, array miscalibration, and partial sensor observations. Beyond producing point estimates, we introduce uncertainty extraction as a key component of remote localization, allowing the system to quantify the reliability of each estimated direction and support confidence-aware fusion across sensors. This approach bridges model-based signal processing and deep learning toward robust, interpretable, and communication-efficient localization in distributed sensing systems. Bio: Raz Zohar MSc candidate @ BGU, supervised by Nir Shlezinger