A document by Maria Elena Valcher. Click on the document to view its contents.

Data scaling is an essential and extensively performed operation in the field of image processing and deep learning. It is used to handle data of varying magnitude and reconstructing tensors in neural networks. While executing such operations in real time, the pre and post processing of data onthe-fly during data transfer reduces memory accesses and delivers ready-to-use data to respective cache, hence playing a vital role in improving the system performance. This paper proposes SCALER, a novel scaling hardware accelerator, aimed at scaling the data in image and NPU Cell [1] formats, on-thefly. It can be configured to scale different types of data-formats, wide range of scaling ratios (both, upscaling and downscaling) and different interpolations namely, Nearest Neighbor, Bilinear Interpolation and Half Pixel Center. This paper also illustrates a scenario where SCALER is integrated inside a DMA which caters to NPU subsystem and supports multiple data transfer paths. SCALER supports on-the-fly processing by overlapping the data fetching and pre-processing time. The current architecture is designed to achieve a maximum throughput of 25.6GBps in real time at 800MHz. Keeping a generic base architecture for multiple data-formats and without any image size restriction, SCALER aims improving performance of DSP/NPU applications and reducing power consumption.

This paper explores enhancements in Adaptive Cruise Control (ACC) using Frequency Modulated Continuous Wave (FMCW) radar to optimize performance in varied environmental clutter conditions, including rain, fog, and insect interference. By implementing Constant False Alarm Rate (CFAR) algorithms, the study reduces clutter impact, enhancing radar accuracy for both range and velocity measurements of multiple targets. Results demonstrate that CFAR integration mitigates errors in velocity and distance estimation, particularly under adverse conditions, supporting ACC reliability over a range of up to 200 meters. This study underscores the effectiveness of advanced clutter rejection in improving safety and functionality in ACC systems for real-world applications.

Myocardial Infarction (MI), commonly referred to as a heart attack, remains a significant challenge in early detection, compounded by the vulnerability of medical data to privacy breaches. This article introduces MI-System, a pioneering framework aimed at facilitating early MI diagnosis through an advanced industrial cyber-physical system (ICPS) coupled with robust privacy-preserving mechanisms. The system is built around a custom-developed Internet of Medical Things (IoMT) device and a state-of-theart security model. MI-System integrates three key layers: 1) the user layer, which gathers photoplethysmography (PPG) signal data via an IoMT device and Smartwatch; 2) the client Layer, where local processing and feature extraction takes place with stringent privacy safeguards; and 3) the cloud layer, which leverages a Federated Learning (FL)based architecture to maintain the integrity of raw data while securing the classification model through differential privacy (DP) and a secure aggregation framework that synchronizes model updates from client devices. To further protect the aggregation process in the FL scheme, an innovative secure, and privacy-focused asynchronous aggregation framework is proposed. In comprehensive realworld trials involving 813 tests across 794 participants, MI-System achieved a remarkably accuracy of high accuracy of 93.51% from the IoMT device, and smartwatches via Android app 94.27%, with minimal latency. This breakthrough enhances MI detection and sets new standards for privacy in smart healthcare, positioning it as a promising candidate for clinical integration worldwide.

The escalating complexity and adaptability of ransomware attacks have exposed critical gaps in traditional detection models, which struggle to effectively identify evolving threat behaviors without a significant margin of error. The Adaptive Pattern Scoring Mechanism (APSM) offers a novel approach to this challenge, leveraging a dynamic scoring framework that recalibrates based on real-time behavioral anomalies, enabling enhanced precision in distinguishing ransomware activities from benign events. APSM addresses inherent limitations in signaturebased and heuristic methods by integrating an adaptive scoring algorithm, which continuously adjusts detection thresholds and assigns differential weights across key behavioral dimensions, achieving a robust balance between detection accuracy and resource efficiency. Empirical evaluation reveals APSM's high accuracy and precision rates, with lower false-positive rates than conventional approaches, indicating its significant capability to detect ransomware with minimal operational disruption. The integration of dynamic scoring not only optimizes APSM's performance in real-time environments but also enables scalability across high-load conditions, marking a substantial advancement in automated cybersecurity infrastructures. Ultimately, APSM provides a sophisticated, resource-efficient solution that aligns closely with the requirements of modern ransomware detection, showing its value as a responsive tool in proactive threat management.

We present a novel AI-based ideation assistant and evaluate it in a user study with a group of innovators. The key contribution of our work is twofold: we propose a method of idea exploration in a constrained domain by means of LLM-supported semantic navigation of problem and solution spaces, and employ novel automated data input filtering to improve generations. We found that semantic exploration is preferred to the traditional prompt-output interactions, measured both in explicit survey rankings, and in terms of innovation assistant engagement, where 2.1x more generations were performed using semantic exploration. We also show that filtering input data with metrics such as relevancy, coherence and human alignment leads to improved generations in the same metrics as well as enhanced quality of experience among innovators.

Modern cars are now much more connected than they were a few years ago, thanks to the rapid development of embedded technology. This had made them more vulnerable to attacks. The Controller Area Network (CAN) bus, a widely used communication standard in automotive systems, plays a crucial role in the interconnection of onboard electronic components. However, the lack of inherent security mechanisms in the CAN bus makes it a prime target for malicious attacks, compromising the system such as the Denial of Service (DoS), fuzzy, spoofing, and replay attacks. In this paper, we propose a machine learningbased intrusion detection system for identifying attacks on in-vehicle CAN bus communication. We train and test Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models on two public datasets (Car-Hacking and CAN-Intrusion) and our self-created dataset, named Bus-CAN-Attack, which was generated using the ICSim simulation tool. Using the selected hyper parameters, we achieve impressive detection accuracy with the fine-tuned models varying from 89% to 99% for the different datasets.

Numerical methods are essential for simulating transient physical phenomena, providing crucial support to engineers and researchers in fields such as defense, space, and the natural sciences. By modeling these phenomena with partial differential equations (PDEs), deeper insights can be gained. Among the various PDE-solving methods, the Crank-Nicolson technique stands out for its efficiency in handling time-dependent PDEs, offering precise and stable solutions across diverse applications. In this work, we present a 3D GPU-accelerated code optimized for parallel execution in three-dimensional space-time domains. The PDE solver is applied to quantum wave function information processing, specifically in the context of the circular double-slit experiment. Additionally, we utilize the solver to simulate a dual Mach-Zehnder interferometer to model the Fenna-Matthews-Olson (FMO) complex, developing a natureinspired algorithm based on its quantum dynamics. This simulation demonstrates near-perfect intelligence in a photosynthetic system without any neural network. This work may help in the design of solar cells, potentially increasing their efficiency to a greater extent.

We introduce a novel deep graphical representation that integrates game theory principles with the laws of statistical physics, enabling feature extraction, and pattern classification within a unified learning framework. In our approach, neurons in a network are analogous to players in a game theory model. Each neuron, viewed as a classical particle governed by the laws of statistical physics, corresponds to a set of actions that represent specific activation values. Neural network layers are conceptualized as sequential cooperative games. The feedforward process in deep learning is interpreted as a sequential game with each game involving multiple players. During training, neurons are evaluated iteratively and filtered based on their contributions to a payoff function, which is quantified using the Shapley value driven by an energy function. Neurons that significantly contribute to the payoff form strong coalitions, and only these neurons are allowed to propagate information to the next layers. The Shapley value facilitates model regularization, thereby improving overall performance. We applied this framework to facial age estimation and gender classification tasks. Experimental results show that our approach outperforms traditional multi-layer perceptron and convolutional neural network models in terms of efficiency and accuracy.

In today’s fast paced world, the healthcare sector and medication management has been facing challenges like long waiting times, understaffing in hospital facilities and pharmacies, and risks to patient safety from incorrect prescriptions. We have witnessed this during the COVID-19 pandemic, and the consequences can be cruel. This research proposes an innovative solution: an Automatic Drug Dispenser System that gives QR-based prescriptions to streamline the process, reduce waiting times, and improve medication tracking. The system enhances patient safety by employing AI-Powered Drug Interaction Warnings. Patients input their medication list into the dispenser, which first runs a rule-based check for known interactions using an interaction database. If no critical issues are found, machine learning algorithms (DeepDDI) perform an additional analysis to detect any potential risks. The system then generates a slip with details on medications, dosages, timings, and any warnings, which the patient receives along with their dispensed medication. This approach aims to revolutionize medication management, improving both efficiency and patient safety.

For the lower microwave frequency bands such as the Ku-band operation, the design of efficient dielectric waveguide antennas is still a major challenge due to several factors such as unavailability of commercial substrates of adequate thickness, bulky feeding networks, and the requirement of large radiating length. The high permittivity substrates have generally been used in the past to reduce the thickness of the radiating structure, but it results into performance degradation due to high dielectric losses, dispersive propagation, and poor impedance matching at the radiating aperture. To overcome these challenges, the concept of a novel substrate integrated travelling wave antenna using the low-index core dielectric waveguide (LICDW) for operation in the Ku-band is proposed in this work. The LICDW structure here consists of a three-layer dielectric waveguide with a relatively lower refractive index core layer sandwiched between two outer cladding layers, which basically confines the electromagnetic energy within the central core enabling lowloss, less-dispersive wave propagation. The proposed antenna is excited by a grounded coplanar waveguide-fed-substrate integrated waveguide (GCPW-fed-SIW) based magneto electric dipole array feeding network to enhance the overall radiation performance. A comprehensive parametric study and the detailed field analysis including numerical simulation is carried out to optimize various parameters of the structure. The optimized structure is fabricated and tested to validate the proposed design methodology. Experimental results confirm the superior performance of the proposed radiating structure compared to the existing Ku-band travelling wave antennas, thereby showcasing its potential for integration into compact, planar RF modules in small satellite systems.

Lakes and reservoirs are essential freshwater sources, yet global monitoring of surface water storage faces challenges due to sparse gauge observations and limitations in remote sensing techniques. The lack of detailed knowledge about spatiotemporal dynamics of inland surface water storage impairs accurate global weather forecasting, Earth system modeling, and water management planning. To address this, we present Water area Tracking from Satellite imagery (WaTSat), a Google Earth Engine-based tool that generates long-term water area time series for global lakes and reservoirs from satellite imagery. Requiring minimal user input, low storage demand, and efficient computational overhead, the tool automates multiple tasks to modify the lake shoreline search area, identify and remove cloud-contaminated images, delineate the lake extent, detect and remove the outliers, and generate the water area time series. The initial version of the WaTSat tool utilizes the MODIS MOD09Q1 product to generate water area time series from February 2000 until the present. Validation against altimetric water level time series for 40 global lakes of varying sizes and regions demonstrates an average correlation of 0.89, confirming WaTSat's capability to accurately estimate surface water area and capture long-term trends and annual fluctuations. The tool's outputs can make a significant contribution to both scientific studies and operational applications in water resource management, hydrological modeling, and climate studies.

This study presents an approach to classify abnormalities in video capsule endoscopy (VCE) frames using a modified ResNet101 architecture with Squeeze and Excitation (SE) blocks. The aim was to build a generalized model capable of automatic abnormality detection across ten classes namely Angioectasia, Bleeding, Erosion, Erythema, Foreign Body, Lymphangiectasia, Polyp, Ulcer, Worms, and Normal. Our approach involves augmenting and balancing the dataset to address class imbalance, followed by training and evaluating the model. The model achieved a mean AUC of 0.984, mean specificity of 0.990, mean average precision of 0.839, mean sensitivity of 0.759, and a balanced accuracy of 0.780. These results demonstrate the potential of SE-ResNet101 for VCE abnormality detection, with opportunities for further improvement in sensitivity and overall accuracy.

Understanding a product's image within a specific market is essential for evaluating and developing effective internal and competitive strategies to enhance product performance. This article discusses the application of Machine Learning (ML) techniques for understanding and proactively predicting a drug's performance rating across multiple dimensions. The ML models developed in this research can analyze shifting patient sentiments toward drugs over time, enabling pharmaceutical organizations to take proactive measures in response to changes in patient sentiment within a given market. Multiple ML algorithms were explored for this purpose, including Support Vector Machine (SVM), Deep Neural Networks, and Bidirectional Encoder Representations from Transformers (BERT). The performance of each algorithm was evaluated on both training and testing datasets sourced from the UCI Machine Learning Repository. The algorithm demonstrating the best performance has been selected for implementation as a front-end application for real-time drug rating prediction based on review text.

The SONICOM / IEEE LAP Challenge is a research competition to push the boundaries of Head-Related Transfer Function (HRTF) analysis and synthesis. Participants are tasked to enhance HRTF dataset accuracy and spatial resolution across various contexts. The 2024 edition of the challenge was divided into two tasks: Task 1 focused on harmonizing HRTFs measured across different labs, while Task 2 involved spatially upsampling sparse HRTF data to a denser grid. The LAP Challenge encouraged participants to develop innovative solutions within a common evaluation framework developed by the organizers. This technical report provides a detailed overview of the submissions, evaluation methodologies, and results for both tasks. All results are fully reproducible using the publicly shared evaluation code, ensuring transparency and allowing for further validation by the research community. It is important to note that future journal publications of the LAP Organizing Team should build upon this data to reflect on the challenge itself, its limitations, and the opportunities it has unveiled, thus expanding on the broader scientific value of this entire experience beyond the competitive aspect.

The emergence of millimeter-wave-based 5G/Beyond 5G and 6G communication technologies has led to advancements in high data rates and low latency. However, there is a massive challenge in implementing the hardware platform including the millimeter-wave antenna. Over-the-air (OTA) testing is crucial for evaluating these antennas, often requiring large facilities to obtain far-field (FF) characteristics. In addition, the measurement accuracy is often limited mainly due to the path loss of the millimeter waves in the far field. Therefore, it is necessary to test the antenna in the near-field (NF) region. In this paper, we present a practical method for obtaining an FF pattern of the antenna-under-test by using an NF-to-FF transformation method with a probe correction technique. The feasibility of the process is validated by the OTA measurement results with a dedicated NF scanning system, which accurately reconstructs the FF patterns from the NF measurement results. Additionally, we have demonstrated that a custom-designed multi-probe can be employed with the transformation method to save measurement time.

Video Frame Interpolation (VFI) involves generating intermediate frames by leveraging both preceding and succeeding frames to enhance video quality. Traditional methods, particularly those based on U-Net architectures, often suffer from high computational complexity and memory demands due to their large parameter counts. In this paper, we introduce the Square Funnel Network (SF-Net), a novel architecture that significantly reduces the number of parameters while maintaining competitive performance. SF-Net employs a unique design that increases the third dimension of the input frame in deeper layers instead of increasing the number of filters, resulting in a simpler and more efficient model. Our architecture utilizes no more than 64 filters in nearly all layers, with only the final two layers using 128 filters each. Through both objective and subjective evaluations, SF-Net demonstrates high visual quality and efficiency, making it suitable for applications with limited computational resources. This work presents a promising direction for VFI, emphasizing parameter reduction without sacrificing performance.

In recent years, deep learning [1] has revolutionized various fields, such as computer vision, natural language processing [2], and speech recognition, by achieving state-of-the-art performance on numerous complex tasks [3]. These achievements are often attributed to the availability of large-scale, labeled datasets and the capacity of deep neural networks to learn rich feature representations. However, one critical challenge remains: the performance of deep learning models significantly deteriorates when there is a domain shift—a mismatch between the data distributions of the training set (source domain) and the test set (target domain).