Applications of Machine Learning 2024

This research investigates the integration of transfer learning techniques with data augmentation strategies to enhance model performance when transitioning across different domains.

This study presents a deep learning-based method for brain tumor detection using MRI scans to improve diagnostic accuracy. Given the grave statistics of brain tumor incidence and mortality, and the low 35.7% survival rate, advanced techniques are crucial. Our research utilized a dataset of 20,670 MRI images to train and test MobileNetV2 and ResNet50 models. ResNet50 significantly outperformed MobileNetV2, achieving 99.1% validation and 93.6% test accuracy, proving its potential for precise classification of normal, glioma, meningioma, and pituitary brain structures. These findings highlight the transformative impact of deep learning in enhancing brain tumor diagnostics.

The UN reports that almost 700 million people can't afford food, while 1.3 billion tons of food are wasted yearly—sufficient to feed the hungry worldwide four times over, as per the Food and Agriculture Organization. Roughly 14% of food waste occurs during transport. To tackle this, the solution utilizes deep learning to predict agricultural produce decay, enabling timely interventions to reduce spoilage and enhance food accessibility and affordability. The development process started with the placement of tomatoes and strawberries in multiple locations with different conditions. Data such as temperature, humidity, and images were collected at intervals until decay. This data trained regression AI models to forecast decay, and detection AI models to isolate produce in large batches. The best models were selected through experimentation with various parameters. These models were then deployed a camera-attached Raspberry Pi prototype, which can be installed during transport, monitor produce, and alert supervisors if any spoilage is detected or predicted to occur soon, so that protective measures can be taken, and in turn prevent food waste and lower prices.

Glaucoma (GC), diabetic retinopathy (DR), and cataracts are major causes of vision loss globally and can be managed effectively with early detection. Traditionally, ophthalmologists manually diagnose diseases using retinal scans which are prone to misdiagnosis, leading to research with artificial intelligence for eye disease diagnosis. Machine learning (ML) statistical algorithms and deep learning (DL) architectures are known to have high performance in medical image analysis. In this study, ResNet50, MobileNetV2, and VGG16 transfer learning architectures were experimented with, along with MLP Classifier, KNN, and Random Forest ML classifiers. A public dataset that had diseased and normal retinal images was used for testing. During the experiments, hyperparameters for each method were modified to determine the best accuracy an algorithm could achieve. The ResNet50 neural network architecture had the highest testing accuracy (91.51%) among the DL methods used and was used during featurization, creating a more accurate model with the MLP at 92.04%. The knowledge from this study has the potential to aid, hasten, and improve eye disease diagnosis for cataracts, GC, and DR.

This project addresses the problem of identifying road damage. Traditional road inspection approaches typically involve scheduled maintenance and repair activities over predetermined periods. We utilized footage taken by ourselves and a public dataset of 6359 images to automate defect detection using our YoloV8 object detection model focusing on road damage. This model, validated with a mAP50 value, shows its effectiveness and ability to run in real-world scenarios. By incorporating new potholes and cracks categories, we enhanced the model’s ability to find specific road defects. This study not only introduces an original and refined dataset but also demonstrates the YOLOV8 algorithm based model’s potential in streamlining road damage detection.

Direct imaging of exoplanets is a challenging task that involves distinguishing faint planetary signals from the overpowering glare of their host stars, often obscured by time-varying stellar noise known as "speckles". The predominant algorithms for speckle noise subtraction employ principal-based point spread function (PSF) fitting techniques to discern planetary signals from stellar speckle noise. We introduce a benchmark developed in the machine learning (ML) framework PyTorch. This work enables ML techniques to utilize extensive PSF libraries to enhance direct imaging post-processing. Such advancements promise to improve the post-processing of high-contrast images from leading-edge astronomical instruments like the James Webb Space Telescope and extreme adaptive optics systems.

Deep learning and image processing play a significant role in the classification and segmentation of breast tumor images, specifically in ultrasound (US) modalities, with the goal of supporting clinical decisions and elevating healthcare quality. Nevertheless, direct utilization of US images poses challenges due to their inherent noise and diverse imaging modalities. In this study, we address these challenges by introducing a three-step image processing scheme. This scheme incorporates speckle noise filtering using a block-matching three-dimensional filtering technique, region of interest highlighting, and RGB fusion. We applied a deep learning model for transfer learning across three datasets: BUSI, Dataset B, and KAIMRC. Employing a fivefold cross-validation on the BUSI and KAIMRC datasets, our model with the proposed preprocessing step achieved improved accuracy rates of 87.8% and 85.2%, respectively. In comparison, its counterpart without the proposed method achieved 80% and 81.6%, respectively. Our method demonstrated significant improvement while maintaining generality across diverse datasets.

As the structure of semiconductor devices becomes complex, it becomes more important to monitor fabrication process. Optical critical dimension (OCD) metrology system is one of the most widely used tool to monitor critical dimenstions (CD) of nanostructures. In the monitoring system, a machine learning model is used to evaluate CDs from measured spectrums. To train the model, synthetic dataset which is several thosand pairs of the simulation structure and spectrum, is used. In this study we suggest a optimization method to generate the database for machine learning for optical monitoring system.

The brachial plexus, a crucial nerve network in the shoulder, facilitates movement and sensory signals to the arms and hands. The brachial plexus block, a primary anesthetic technique for upper limb surgeries, presents challenges in nerve identification and localization, often leading to potential complications. This abstract introduces a solution through the application of convolutional neural networks (CNNs), specifically employing a U-net model trained on the Ultrasound Nerve Segmentation competition dataset. The model effectively automates the identification and segmentation of the brachial plexus, achieving a dice coefficient (DSC) of 0.86, indicating high accuracy. This method promises to enhance the precision of echo-directed brachial plexus block procedures, aligning with advancements in deep learning to improve surgical outcomes.

Vector-quantized variational auto-encoders (VQ-VAEs) were trained on labelled synthetic quadrature signals in an effort towards data-driven RF-channel estimation. Using simulated GNU Radio datasets from RadioML, short samples of IQ data with various digital modulations and SNRs were used to train a conditional VQ-VAE and a PixelSnail network was utilized to learn and generate codes. The VQVAE-generated samples show promise of reproducing the GNU Radio simulated channel effects when compared to the training data, SNR calculations, and modulation recognition classifier results.

The presence of broadleaf and grassy weeds in fallowed fields threatens crop yield and quality, as they compete with crops for essential resources. Traditional weed control methods, relying on broad-spectrum herbicides, are both environmentally damaging and costly. Vision Transformers (ViT) have emerged as a revolutionary tool in Computer Vision, offering advanced capabilities in image analysis. This study utilizes ViT, specifically YOLOS (You Only Look One Sequence), to automatically detect and classify broadleaf and grassy weeds in fallowed fields. With a dataset of 15,542 real field images, the model achieved a precision of 90.7% and an average recall of 86.3%. The results demonstrate YOLOS as a promising alternative for accurately distinguishing between broadleaf and grassy weeds in fallowed fields.

Hyperspectral sensing is a valuable tool for detecting anomalies and distinguishing between materials in a scene. Hyperspectral anomaly detection (HS-AD) helps characterize the captured scenes and separates them into anomaly and background classes. It is vital in agriculture, environment, and military applications such as RSTA (reconnaissance, surveillance, and target acquisition) missions. We previously designed an equal voting ensemble of hyperspectral unmixing and three unsupervised HS-AD algorithms. We later utilized a supervised classifier to determine the weights of a voting ensemble, creating a hybrid of heterogeneous unsupervised HS-AD algorithms with a supervised classifier in a model stacking, which improved detection accuracy. However, supervised classification methods usually fail to detect novel or unknown patterns that substantially deviate from those seen previously. In this work, we evaluate our technique, as well as other supervised and unsupervised methods, using general hyperspectral data to provide new insights.

We utilize Stimulated Raman spectroscopy with deuterium oxide (DO-SRS) to acquire hyperspectral images of mouse hippocampal tissue. With DO-SRS, we can identify carbon-deuterium (C-D) bonds that are indicative of biomolecule synthesis, providing metabolic information of the tissue. Through k-means clustering analysis of the Raman C-H stretching and C-D bands, we can distinguish the mouse hippocampal regions based on the vibrational mode of deuterated proteins and lipids.

In the field of glioblastoma radiomics, precise tumor characterization poses a significant hurdle. Glioblastoma, known for its high severity and common occurrence, demands sophisticated diagnostic methods for efficient treatment. Radiomics emerges as a key technique in this domain by leveraging the quantitative analysis of medical images. In this study, different feature selection methods and machine learning classifiers designed for glioblastoma radiomics are carefully compared and contrasted. It systematically reviews and contrasts the effectiveness of different feature selection processes, such as Random Forest feature importance, correlation-based feature selection, mutual information, L1 regularization (Lasso), and XGBoost feature importance. The study looks at the performance of several classifiers at the same time, such as K-Nearest Neighbor, Quadratic Discriminant Analysis, and Multilayer Perceptron. The primary goal is to discern the most potent combinations that enhance predictive accuracy and reliability, thereby making a substantial contribution to the development of personalized medical treatments for brain tumors.

Explainable Machine Learning (XML) approaches are crucial for medical information processing tasks, particularly for multi-omics data analytics. The XML system not only provides better performance but also explains the inside of the finding better. Here, we proposed an end-to-end explainable system for analyzing high dimensional RNA-seq data using an unsupervised gene selection approach and supervised methods, including Deep Neural Network (DNN), Deep Convolutional Neural Network (DCNN), Support Vector Machine (SVM), and Random Forest (RF). The proposed approaches evaluate with publicly available datasets for five different cancers and Kawasaki disease (KD) classification. The deep learning-based approaches yield the 99.62% and 99. 25% average testing accuracy for cancer and KD classification tasks. Additionally, we introduce an explainable system that demonstrates the ability to select cancer and disease-specific gene sets, which could be used for further analysis to discover the biological inside of the cancers and KD diseases.

Lung segmentation supports essential functions in the realm of computer-aided detection via chest radiographs (CRs). In this research, we present an approach of segmenting the lungs by separating them down the spinal column and having separate networks for right and left lung respectively. Using DeepLabV3+ as the fundamental deep-learning framework, results are presented for publicly available datasets such as Japanese Radiological Scientific Technology (JRST) and the Shenzhen datasets. We achieve an overall accuracy of 98.8% and an IoU (Intersection over Union) of 0.977 for a set of 100 test cases thereby setting a new benchmark.

This study presents a novel Convolutional Neural Network (CNN) architecture for classifying axial curvature maps of the anterior corneal surface into five distinct grades of keratoconus severity. Leveraging a meticulously labeled dataset of 3,832 corneal maps, the model achieved an overall recall of 78.5% and an average precision of 76.9%, demonstrating notable proficiency in distinguishing between normal eyes and various stages of keratoconus.

Breast cancer research is prominent, constituting 25% of female cancer cases, impacting 1.7 million yearly. While biopsies are the standard, the surge in cases challenges manual assessment of histological images. Optical traits like polarization offer potential for early detection. Automated methods expedite diagnoses, eliminating specialized expertise need. This study aims to provide a rapid alternative to traditional histology using full-field polarization sensitive optical coherence tomography (FF-PS-OCT). The automated FF-PS-OCT system demonstrates performance akin to conventional histology, achieving remarkable classification results. The technique's rapid and label-free imaging advantages are pivotal for well-informed medical decisions.

Missing data presents a formidable obstacle in asteroid databases, directly affecting the accuracy of taxonomic classification. This study undertakes an evaluation of machine learning methodologies to address this challenge and enhance taxonomic classification. The techniques under scrutiny encompass logistic regression, support vector machines, AdaBoost, XGBoost, multilayer perceptrons, and K-nearest neighbors. By harnessing a comprehensive dataset, our investigation seeks to pinpoint the most efficacious imputation method, thereby bolstering the reliability of asteroid databases and deepening our comprehension of asteroid classification.

Characterizing rapid changes in sea ice is crucial for Arctic routing. Towards this end, we previously developed a probabilistic, polarimetric approach to obtain fine-resolution ice/water labels (hard labels) and uncertainties (soft labels) in complex-format Sentinel-1 (S1) synthetic aperture radar (SAR) imagery. Here, we assess the benefits of such probabilistic labels as well as the importance of incorporating model probability calibration by training machine learning models of varying complexity with S1 intensity-format products as input. We find that training with soft probabilities adds value as model complexity increases, and in the absence of soft labels, incorporating probability calibration is beneficial.

Deep learning and machine learning systems have proven to be effective for detecting various hazards/disasters. In this research, we present an automated deep learning based algorithm to accurately segment and identify various regions affected by fire using videos captured through drone. In addition, we also identify the presence of smoke within each frame. We annotate a drone imagery dataset to segment burned, burning, and unburnt regions within a frame. Based on a set of 24666 images, our segmentation algorithm achieves a mean IOU of 0.93 and our classification algorithm for identifying smoke achieves an overall accuracy of 92%.

Spatio-temporal predictive learning, aimed at forecasting frames with self-supervision, adapts to both traditional and all-sky camera images, the latter requiring considerations of distortion and symmetry. Despite progress, challenges in standardization and reproducibility persist. Comprehensive evaluations across various datasets highlight the need for specialized models for all-sky imagery, revealing significant performance differences.

This research addresses the challenge of camera shake in agricultural image-based phenotyping caused by vibrations from motor-driven mechanical arms. It proposes a novel solution integrating Sliding Mode Control (SMC) with Artificial Intelligence (AI) for effective vibration control in agricultural manipulators. Through rigorous Lyapunov analysis, the stability of this control system is validated. An innovative addition is the use of an active actuator for torsional vibration mitigation. Compared to traditional PD/PID controllers, this new approach demonstrates superior vibration attenuation, bridging a significant gap in agricultural practices with a cost-effective and efficient solution for improved imaging precision.

This study tackles agricultural product decay, a major contributor to food waste. The research focuses on early biological activity identification, particularly in the aging and decaying process, to enhance quality control and predictive food safety monitoring. Introducing an innovative spatial-temporal speckle correlation technique based on machine learning, the study highlights the potential for proactive measures to prevent food waste. Employing low-cost, rapid, and non-destructive speckle metrology, the experiment utilizes a coherent laser to illuminate principal fruits, tracking dynamic speckle patterns over time. Fluctuations in these patterns, influenced by parameters like drying, aging, and ripening, are assessed using a 2D cross-correlation method coupled with machine learning. Additionally, a procedure for quality and contrast testing is outlined to overcome challenges associated with low-quality images in machine learning. The findings demonstrate that dynamic speckle activity mirrors the physical and biochemical changes in the fruit's shelf life. The study proposes a valuable approach for reducing food waste through early decay detection and quality assessment.

We have studied the ResNet50 with Grad-CAM and proposed an approach to enhance their explainability. We have adapted the higher-order statistics of feature maps since they can retain the semantic meanings of the feature maps. The goal of the approach is to compare the feature maps of the final convolutional layer of ResNet50 for an image and its Grad-CAM version. We have conducted a simulation with the images of Cardinals and American Goldfinch to retrain the ResNet50 model and develop ResNet50new to classify the two types of birds with explainability. Simulations show that the use of higher-order statistics allows the development of well-defined k-means clusters of the ResNet50 feature maps (associated with the images and their Grad-CAM versions) and the enhancement of explainability.

The overfitting of deep learning (DL) models is one of the problems in artificial intelligence (AI) systems. It occurs when a DL model learns not only the patterns inherent to the data but also learns the noise characteristics and random fluctuations in the data. The identification of the feature maps that support/oppose AI’s decisions under uncertainty, caused by the noise characteristics and random fluctuations in the data, is challenging. This paper studies and proposes a Bayesian search theory-based approach that assumes prior knowledge of the feature maps of the noise-free observations and generates posterior probabilities to find correlated feature maps of the noisy observations. Simulations with VGG16, bird images, and noise degradation show that we can precisely locate the feature maps and their semantic meanings that support/oppose the decision (i.e., the prediction) of an AI system under uncertainty.

The ongoing paradigm shift in healthcare towards personalized and precision medicine is posing a critical need for noninvasive imaging technology that can provide quantitative tissue and molecular information. Magnetic resonance signals from biological systems contain information from multiple molecules and multiple physical/biological processes (e.g., T1 relaxation, T2 relation, diffusion, perfusion, etc.). So, magnetic resonance imaging (MRI) is inherently a high-dimensional imaging technology that can acquire structural, functional and molecular information simultaneously. In practice, due to the curse of dimensionality, MRI experiments are often done in a low-dimensional setting to acquire biomarkers one at a time. Such a “divide-and-conquer” approach not only reduces data acquisition efficiency but also makes it difficult to obtain molecular information in high resolution. By synergistically integrating machine learning with sparse sampling, constrained image reconstruction and quantum simulation, we have successfully demonstrated ultrafast high-dimensional imaging of the brain. This talk will give an overview of this unprecedented omni imaging technology and show some exciting experimental results of brain function and diseases.

With 140+ petabytes of historical data holdings, 3.8 million square kilometers of daily multi-spectral collection, integration of Synthetic Aperture Radar and newly launching assets every quarter, the opportunities to develop insight from sense making technologies at Maxar are ever growing. During this discussion, we will cover the challenges of collecting, organizing, and exploiting multi source electro-optical remote sensing systems at scale using modern machine learning architectures and techniques to derive actionable insights.

Robustness to image quality degradations is critical for developing convolutional neural networks (CNNs) for real-world image classification. This paper delves into previous results of how optical aberrations and optical scatter degrade classification performance and explores how they cause classification errors to manifest within CNN layers.

IC SEM image analysis for trojan detection in real-time is an essential task for rapid design verification. Deep learning-based methods face challenges, demanding a significant number of high-quality data, and yielding a high-data acquisition time to build up the system. This study focuses on deep adversarial learning based on low to high-resolution SEM image mapping to address the data challenge. Essentially, the method employs a few-shot learning incorporated cycle-generative adversarial network to reduce the model’s dependency on many high-quality data. Leveraging few-shot learning-based low-to-high image mapping mitigates the high acquisition time problem, and potentially opens the possibility of implementing a real-time trojan detection system for design verification.

This article outlines the implementation of CNN models designed to enhance the efficiency of manually validating engineered projects. Our approach involves utilizing computer-aided design simulation image captures as training data for our pipeline. We integrate a real-time color-filtering and fiducial scaling normalization process on any constructed project images for our algorithm to perceive them in a consistent manner with simulation images from the model training.

We use machine learning to enhance the resolution of local near-field probing (LNFP) measurements, specifically when the probe size exceeds the dimensions of the device under examination. The study employs a simple numerical model mimicking LNFP setups and evaluates two machine learning methods—linear regression and fully connected neural networks (FCNNs). Results indicate that both methods can accurately predict the electric field distribution along a photodetector, achieving a spatial resolution of one-tenth of the wavelength of the excitation with minimal relative errors. Furthermore, the study suggests that linear regression is efficient for extensive training datasets, while FCNNs outperform in accuracy for smaller datasets. The findings have broader implications for improving resolution in various measurement setups.

This report describes our study where we implemented Convolutional Neural Networks (CNNs), a subclass of ML, to analyze defects in 3D-printed steel samples and move toward a more efficient vision-based metrology technique for metal additive manufacturing. We focused on Selective Laser Melting (SLM) as the primary manufacturing technique. Using a library of SEM images for 3D-printed steel samples allowed us to train our CNN algorithm to recognize and classify the main types of metal printing defects based on their geometry and composition. The primary defects targeted are lack of fusion, balling, and impurities. Following the training of the ML model, we employed the trained CNN algorithm to characterize newly SLM-printed stainless steels. The confusion matrices were applied to break down the trained model's predictions and evaluate its accuracy. The results highlight the promising application of machine learning in the automated metrology of AM products and fine-tuning of AM parameters.

This study addresses the challenge of predicting fatigue cracks, a major cause of structural failure, by proposing a novel methodology combining computer vision with an RNN-LSTM architecture. Traditional methods often fail due to the dynamic nature of crack propagation. The approach transforms image prediction into sequential data analysis, using innovative techniques like skeletonization. Model tuning and parameter fitting are performed, with accuracy assessed against ground truth. Different cross-validation methods yield varying optimal configurations. Multi-image training evaluation shows challenges in developing general models. While effective, predicting irregular crack trajectories remains difficult. Overall, the study presents a promising tool for nondestructive evaluation and repair planning.

Detection of defects and damages due to aging and transient events are important contributors to pipeline accidents and monitoring them together is challenging. We demonstrate an intelligent fiber-optic acoustic sensor system for pipeline monitoring that enables real-time recognition, and classification of defects and transient threats together by analyzing the combined acoustic ND data from the ultrasonic guidedwaves (UGW) and acoustic emission (AE) methods. A 6"carbon-steel pipeline (16-ft long, SCH40) is set up with multiple defects (weld and corrosion) which are monitored by the UGW and interferometric fiber sensors, while transient events (intrusion and impact) are manually generated and are detected using distributed acoustic sensor (DAS). Finally, the convolutional neural network (CNN) is applied onto the acoustic ND data to realize an accurate and automated pipeline health monitoring solution.

Maintaining gas pipeline integrity is paramount for operational safety and reliability. Detecting both physics defects (e.g., corrosion, welding) and non-physical events (e.g., digging, encroachment) is critical. However, replicating physical defects experimentally is challenging, necessitating simulation-based approaches. In contrast, non-physical events can be observed through laboratory and field tests. This study proposes an integrated framework that combines simulated data of physical defects with experimental data of non-physical events. Deep learning classification algorithms are applied to this combined dataset, enhancing detection accuracy. The framework bridges physics-based simulations and data-driven techniques, offering a comprehensive solution for pipeline monitoring. This approach strengthens surveillance systems, bolstering gas distribution network safety and resilience.

The effectiveness of deep learning models in PCB X-ray inspection depends on annotated data. Synthetic data augmentation can improve performance and reduce manual annotation burdens. The Synthetic Data Tuner framework optimizes deep learning models for PCB X-ray inspection. It combines advanced architectures with generative adversarial networks and variational auto-encoders to assess the impact of synthetic data on accuracy, robustness, and generalization. The findings show the importance of balancing synthetic data quantity and performance to maximize improvements without overfitting. This approach advances PCB X-ray inspection and contributes to the computer vision and industrial inspection domains.

This study investigates the enhancement of computer vision resilience in intelligent PCB inspection through advanced adversarial example filtration. PCBs are crucial in electronics and require reliable inspection. Current computer vision models are susceptible to adversarial attacks, which compromise their accuracy. Our approach combines deep learning with adversarial training, allowing the model to adapt to potential threats. A refined filtration mechanism mitigates the impact of adversarial attacks during real-time inspections, yielding promising results. This research emphasizes the ongoing necessity of investigating adversarial example filtration to strengthen intelligent inspection systems.

Segmenting printed circuit board (PCB) components from X-ray images constitutes an essential task for design extraction and reverse engineering. Traditional pre-trained deep learning models face challenges, demanding significant resources and yielding suboptimal results, especially when labeled data of the PCB X-ray is scarce. The Segment Anything Model (SAM) is versatile but encounters difficulties with complex PCB X-ray designs, making accurate segmentation of various components challenging. This study proposes a customized approach, modifying SAM with parameter-efficient fine-tuning and few-shot strategies, addressing challenges like intricate design and X-ray artifacts of PCBs. The methodology focuses on adapting the foundation model efficiently to the unique features of PCB X-ray images. Leveraging few-shot learning mitigates the limited annotated data challenge, potentially offering a novel solution for implementing deep learning in a constrained dataset while capitalizing on a foundation model's capabilities.

Identifying long term user interests, i.e., “evergreen missions”, in retail and e-commerce is a challenging yet important problem. In this work, we propose a machine learning system that is able to identify a user’s long term arbitrary interests by leveraging their purchasing and site interaction history. Our contribution is a system that is composed of three components (1) projecting our listing inventory to an embedding space with a combination of supervised/unsupervised modeling, (2) inferring personalized interests from the embedding space to a user base with attributed interactions, and (3) estimating the repeat interaction rate with inventory through a rigorous statistical approach. Additionally, we provide novel insights by leveraging the supervised neural network model to produce a clustering approach for interest discovery. The approach has been implemented, validated, and rigorously A/B experimented with and is currently in production at Etsy, Inc., powering its several modules.

E-commerce recommendations commonly leverage buyer’s recent activity, which has previously shown to improve sitewide engagement and merchandising sales. While important to exploit what we know about users, long-term growth and a buyer’s ability to discover new products is equally important, albeit underexplored. In this work, we propose methodology to balance personalization with exploration in the context of seasonal and trend-based recommendations, taking advantage of temporal and buyer/seller localization signals. Our approach accounts for an understanding of buyer intent to balance relevancy with serendipity in item recommendations. On our platform, our proposed approach in seasonal and trending recommendations can not only boost immediate buyer experience in company performance but also play a key role in the ongoing growth of the marketplace.

5G New Radio (NR) and communication systems beyond 5G (NextG) are susceptible to cyberattacks. In a cyberattack, an adversarial transmitter will attempt to intercept a legitimate signal or imitate the legitimate signal and transmit forced messages to a receiver. A novel method to identify and locate MIMO rogue base stations (BS) transmitting signals using 5G NR and 6G communication protocols to user equipment (UE) is proposed. A machine learning (ML) approach for distributed and networked set of passive MIMO antennas is proposed to observe the spatial characteristics of the communication channel and detect the existence of rogue base stations. The channel observance from each passive MIMO antenna is fused together and passed to a ML-based decision algorithm to determine whether the packet of information is indeed sent by the legitimate transmitter or by the spoofing BS.

In model compression, filter pruning stands out as a pivotal technique. Its significance becomes particularly crucial as the present deep learning models are developed into larger and more complicated architectures, which leads to massive parameters and high floating-point operations per second (FLOPs). Challenges have appeared due to the high computational demands associated with these advanced model structures. In this work, we introduce two novel methods aimed at addressing the challenges above: innovative automatic filter pruning methods via semi-supervised multi-task learning (SSMTL) hypernetwork and partial weight training hypernetwork, respectively. Both methods effectively train the hypernetwork and enhance the precision of the neural architecture search with reinforcement learning. Compared to other filter pruning methods, our approach achieves higher model accuracy at similar pruning ratios.

The diversity and quality characteristics of a Generative Artificial Intelligence (Generative AI) to generate image outputs are studied by generating random noise vectors using several statistical distributions (e.g., Beta, Cauchy, Gamma, Gaussian, and Laplace distributions) that allow access to the diversified latent space of the images by BigGAN. Simulations using the BigGAN Generative AI show significant diversity can be achieved by drawing random noise vectors from multiple statistical distributions. Simulations also show that the shape deviations of statistical distributions, considering Gaussian as the reference distribution, influences randomness of the noise vectors associated with the latent space of the images.

When laser light beams are allowed to propagate from one end of a multi-mode optical fiber to the other end and further to be output onto a screen, irregular pattern called as speckle patterns can be observed in an output light spot. The authors previously reported the rotating phenomenon of such speckle patterns when the optical fiber was placed onto a support plate in a loop-shape and the support plate was tilted. In this paper, we examined the method for estimating tilted angles to classify speckle pattern images by ResNet-18 trained using transfer learning. As a result, the network with a classification accuracy of approximately 95% in the measurement range of -10 to +10 degrees of tilt angle was realized

In autonomous driving, deep learning models developed without addressing all the corner cases, and unexpected and unknown life-threatening situations, fail in the real world. Models for corner cases in the visible domain are developed on the assumption that objects are visible, but these corner cases can also occur at night or in foggy weather. Visible cameras are not able to give a healthy input to the model in such situations. However, infrared cameras can provide reliable input to detection models. In this study, we propose a synthetic dataset using stable diffusion, which allows the detection of corner cases that may occur in poor light conditions. We train a deep learning model using this dataset and we present a baseline on this dataset.

Modern deep learning methodologies have shown a clear potential for automatically detecting weapons in CCTV images. The most advanced methodologies not only consider the appearance of the weapon but also analyze the individual's body pose, crucial in scenarios with poor lighting and distant cameras. Unlike previous methods that solely rely on simple body poses, this study explores the effectiveness of incorporating hand poses to enhance weapon discrimination. The research highlights the potential of hand poses in distinguishing weapons from other objects and emphasizes that modern surveillance cameras with sufficient resolution and/or zoom can interpret hand poses. The experiments conducted demonstrate the system's potential, comparing it to previous approaches based solely on body pose analysis. The text also addresses practical considerations and limitations when implementing the proposed system with surveillance PTZ cameras.

When an intense laser interacts with a gas of atoms, high-order harmonics are generated. In the time domain, this radiation forms a train of extremely short light pulses, of the order of 100 attoseconds. Attosecond pulses allow the study of the dynamics of electrons in atoms and molecules, using pump-probe techniques. This presentation will highlight some of the key steps of the field of attosecond science.