For synchrotron based user experiments, radiation characteristics such as energy resolution, flux and beamsize are of great importance in terms of the feasibility of the experiment. The main difficulty stems from adjusting the optics and other components throughout the synchrotron beamline to obtain the required energy resolution, flux and beam position values. The optimum parameters are determined by manually (remotely) controlling the each element constituting a beamline, which takes several days even weeks. The main goal of the Genetic Algorithms-based Synchrotron radiation Optimization for an X-ray beamLINE (GASOLINE) project was to optimize the synchrotron beamline components operating in the hard and soft X-ray region through Genetic Algorithms (GA) in order to achieve the required synchrotron beam characteristics for dedicated user experiments. Two other optimization algorithms were employed in addition to Multiple Objective Evolutionary Algorithms and variants (e.g. GA, NSGAII/III etc.), namely Particle Swarm Optimization (PSO) and Artificial Bee Colony Optimization (ABC). Results for the optimization processes designed and implemented in the project show that the bio-inspired approaches employed in the project have achieved significant success in the order of seconds execution time.
Since accelerator based light sources are cutting-edge-technology tools in natural sciences research of the 21st Century, they provide unique user experiments by superior radiation characteristics. On account of this, rapid & high-precision alignment of photon beamline components (including design and optimization of the undulator as a matter of course), is a big concern for many beamline scientists. In this respect, a hybrid in-vacuum X-ray undulator driven by a 6 GeV synchrotron, is optimized through Evolutionary Algorithms (EAs). It is shown that the EA results provide promising design & optimization estimations for photon beamline scientists.
X-ray absorption spectroscopy is an established technique in the application of synchrotron radiation with applications in many scientific fields. Some scientists are not really familiar with the applications of X-rays in science, others have rather complicated sample environments, which require their full attention. Everybody will profit from an automation of the instrument with control-mechanisms for best beam quality for the experiment, automated sample alignment and control, and preliminary data analysis. These automation procedures can benefit from AI-based algorithms.
I will also briefly present an example of data analysis, which is based on an artificial neural network.
Research on the application of small-angle x-ray scattering (SAXS) method, using x-ray free-electron laser (XFEL) images, utilizes normalizing flows for the inversion of experimental X-ray scattering images. One of the main challenges lies in the inversion of such experimental scattering images, which contain various artifacts such as parasitic scattering, slit scattering, beamstop, and detector background. These artifacts pose a significant domain shift for the neural network used in the inversion process. Parasitic scattering typically appears as a Gaussian-shaped cluster around the primary beam, accompanied by scattered photons in the vicinity. Slit scattering manifests as streaks around the primary beam, while the beamstop obstructs the main beam entirely, resulting in a lack of signal. The detector background refers to an offset with some underlying structure. Currently, the simulated dataset is being modified to incorporate these artifacts. However, this approach may not be sustainable in the future, as the exact characteristics of the artifacts are unknown in advance, and there is limited time to model them during the experiment. Hence, the intention is to collaborate in developing a resilient feature extractor capable of extracting features from both simulated and experimental data, even in the presence of unknown artifacts. These extracted features will subsequently be utilized for the inference process in the primary neural network responsible for inversion.
We aim to address the problem by using contemporary deep learning techniques. At present, we are exploring two potential approaches: one involves learning representation through $\beta$-VAE, while the other entails utilizing image-to-image translation methods such as CycleGAN and pix2pix.
Analyzing a significant amount of remote sensing data obtained through Unmanned Aerial Vehicles (UAVs) and satellite-based technologies for Precision Agriculture (PA) applications involves artificial intelligence (AI) methods. AI applications aim to enhance crop yield efficiently while minimizing input costs and environmental impacts in optimal locations. Designing the most suitable deep learning architecture, considering the requirements of PA, is crucial. This architecture should maximize the utilization of multispectral band images found in remote sensing data, offering significant information. However, specific AI techniques, particularly Convolutional Neural Networks (CNNs), experience delays in inference. There is limited research on AI algorithms that integrate into UAVs for real-time PA applications. Therefore, the designed CNN architecture should aim to reduce inference time and establish a balance between speed, memory, and accuracy. This study evaluates the performance of various CNN architectures, considering all PA application requirements for detecting objects such as trees, crops, and wheat yellow-rust disease. The final discussion covers modifications relevant to achieving an optimal CNN architectural design for PA.
Coffee Break
As life expectancy rises, the elderly care sector faces a looming workforce shortage. Ambient Assisted Living (AAL) systems emerge as a promising solution, with Human Activity Recognition (HAR) as a key component. HAR's ability to automatically track activities empowers comprehensive health and well-being assessments, easing the burden on healthcare professionals. Lifelogging and home diary applications further enhance this value by providing valuable insights into daily routines.
Complex activities, however, pose a unique challenge due to their intricate semantic nature. To address this, a machine learning framework that utilizes a single accelerometer sensor to accurately predict complex human activities is proposed. The Locally Weighted Random Forest (LWRF) algorithm, with its hybrid structure and local weighting approach, proves remarkably effective, achieving 91% accuracy for HAR and 91.3% for gender recognition.
This work paves the way for embedding the proposed framework into lifelogging and home diary applications, enabling real-time monitoring of mental status and overall well-being among the elderly. As AAL systems continue to advance, their potential to transform elderly care is undeniable.
Deep learning (DL) has become widely used in many optical sensor scenarios during the last few years. Notable progress has been made in integrating deep learning algorithms, which improve accuracy and reduce noise in optical sensor data. Optical sensors are a highly promising technology for modern intelligent sensing platforms. They are used in many different fields, including defence, security, process monitoring, quality prediction, pollution control, and many others. Although optical sensors have many applications, they also face many difficulties. These difficulties include the need to create large datasets and the resulting slow processing speeds. On top of this, these sensor technologies are very expensive. To overcome these obstacles, deep learning systems and optical sensor technologies must be strategically integrated. This work presents a systematic review of recent works that have effectively used DL algorithms in optical sensor applications. It not only sheds light on the state of these integrations today but also highlights some encouraging avenues for DL algorithms’ continued development in the context of optical sensor applications. Investigating these integrated systems shows how they can be used to get around the problems of processing speed constraints and big datasets, which will ultimately lead to more effective and affordable optical sensor solutions. This work also acts as a lighthouse for future research projects, pointing the way towards fresh directions and offering opportunities for creativity in the ongoing advancement of DL applications in the field of optical sensors. Overall, this work lays the groundwork for future research and development of AI-enabled optical sensor technologies.
Soft X-ray spectroscopy is a valuable tool for extracting information about the electronic properties of atoms and molecules within a given sample. This technique provides insights into various aspects,
including electronic configuration, bonding traits, and the surrounding chemical environment. Nevertheless, a significant portion of pertinent chemical processes, particularly those of a biochemical
nature, transpire within liquid environments.
This circumstance introduces an inherent challenge for experimental endeavors. Consequently, the exploration of molecular systems within their native liquid state necessitates the utilization of advanced
experimental setups and sample delivery mechanisms.
In this presentation, I will delve into the most recent findings derived from compact photon sources that can be operated on a tabletop. These sources harness powerful femtosecond lasers to drive soft
X-ray (SXR) generation, facilitating the practice of time-resolved X-ray absorption spectroscopy (TRXAS).
Through this approach, it becomes feasible to monitor dynamic structural and electronic changes in real time.
Cotton whitefly is an important pest that damages crops and reduces crop yield. An IoT-based data Aggregation methodology is proposed to address the challenges of cotton whitefly pest monitoring and control. Uses sensors and IoT devices provide real-time monitoring and complete control of whitefly pests, while also providing timely information to farmers and supporting decision-making by agronomists. A method of data aggregation has been used to determine the number of pests of cotton whiteflies and the damage caused by them. The data aggregation system is designed to support real-time pest monitoring and control, which is critical for effective pest management and crop protection. This study has demonstrated that IoT-based data collection method has proved useful in detecting and controlling cotton whitefly pests and providing timely information to farmers
Multilabel classification is a type of classification task in which an instance can be assigned many labels at the same time. Multilabed data can be correctly classified using machine learning and deep learning models. Transfer learning is a machine learning method in which a model trained for specific task is used for different but similar task. In this study, we utilized the transfer learning to overcome the multilabel classification problem. As a case study movie poster image classification is selected because in general, movies have multiple labels/genres. For this purpose, movie images from The Internet Movie Database (IMDB) have been downloaded and used. The popular modern pretrained models have been employed for transfer learning. The iterative stratification technique has been utilized to partition the dataset. Each model has been trained and fine-tuned using this dataset. The performance of each model has been compared taking into account metrics such as AUC, f1-score, precision, loss, and hamming loss.
Coffee Break
Femoral fractures, a growing threat to the elderly, are predicted to double by 2050. Early diagnosis and intervention are crucial for joint protection, quality of life, and mobility after surgery. However, misdiagnosis triggers a cascade of delayed treatment, prolonged recovery, and spiralling costs, impacting both patients and healthcare systems.
Deep learning is revolutionizing medical image processing, transforming diagnosis and treatment by extracting hidden patterns from medical scans. Its precision in fracture detection minimizes misdiagnosis, reducing unnecessary procedures and hospital stays, leading to significant cost savings for patients and healthcare systems.
The success of a deep learning architecture is tightly dependent on the values of its hyperparameters. It is possible to increase the performance of a current model by selecting hyperparameters with a genetic algorithm. In our study, we achieved a 1.6% performance boost, in terms of accuracy, for an existing CNN model using a genetic algorithm.
Lunch
The alignment of each analyzer position in X-ray Raman Scattering (XRS) setup requires a great deal of time, therefore using the optimization approaches is necessary. In this study, Non-dominating Sorting Genetic Algorithm-II (NSGA-II) was used on 101 images from detectors while positioning each analyzer. The images first filtered through 2D Gaussian filter to obtain a cloud of dots representing the spread of the X-ray, then spot size and flux values were optimized first separately and later simultaneously. Flux is related to the amplitude of the Gaussian distribution of the spread of the rays, whereas spot size is related to the sigma values on the vertical axis x and the horizontal axis y. The first results show that minimum sigma-x, sigma-y values and maximum amplitude are provided at different positions, thus NSGA-II is run on multi-objective mode. Pareto set is found with six elements.
In recent times, visual surveillance systems have undergone rapid advancements, exhibiting increased capabilities and widespread integration of artificial intelligence. Concurrently, these surveillance systems have raised concerns by exposing the public to emerging privacy and security risks. Instances of overt misuse of surveillance technologies have surged, prompting the implementation of data privacy regulations such as GDPR in Europe to establish guidelines for responsible data collection and processing.
Despite these regulatory measures, there remains a pressing need for a secure and private approach to train sophisticated machine learning and deep learning algorithms. In this paper, we propose a method that prioritizes privacy in visual surveillance. Initially, we curate a dataset consisting of videos with preserved privacy. The content within these videos is obfuscated using a combination of Gaussian Mixture Model (GMM) and selective encryption techniques. Subsequently, we employ this privacy-preserved dataset to train high-performance object detection models.