Abstract
Data stream mining seeks to extract useful information from quickly-arriving, infinitely-sized and evolving data streams. Although these challenges have been addressed throughout the literature, none of them can be considered "solved." We contribute to closing this gap for the task of data stream clustering by proposing two modifications to the well-known ClusTree data stream clustering algorithm: pruning unused branches and detecting concept drift. Our experimental results show the difficulty in tackling these aspects of data stream mining and the sensitivity of stream mining algorithms to parameter values. We conclude that further research is required to better equip stream learners for the data stream clustering task.

Abstract

Abstract
Diabetic Retinopathy (DR) is a leading cause of blindness in working age adults. DR lesions can be challenging to identify in fundus images, and automatic DR detection systems can offer strong clinical value. Of the publicly available labeled datasets for DR, the Indian Diabetic Retinopathy Image Dataset (IDRiD) presents retinal fundus images with pixel-level annotations of four distinct lesions: microaneurysms, hemorrhages, soft exudates and hard exudates. We utilize the HEDNet edge detector to solve a semantic segmentation task on this dataset, and then propose an end-to-end system for pixel-level segmentation of DR lesions by incorporating HEDNet into a Conditional Generative Adversarial Network (cGAN). We design a loss function that adds adversarial loss to segmentation loss. Our experiments show that the addition of the adversarial loss improves the lesion segmentation performance over the baseline.

Abstract
In order to enhance the scientific exploitation of adaptive optics (AO)-assisted observations, we investigate a novel hybrid concept to improve the parametric estimation of point spread function (PSF) called PSF Reconstruction and Identification for Multiple-source characterization Enhancement (PRIME). PRIME uses both focal and pupil-plane measurements to estimate jointly the model parameters related to the atmosphere [Cn2(h), seeing] and the AO system (e.g. optical gains and residual low-order errors). Photometry and astrometry are provided as by-products. The parametric model in use is flexible enough to be scaled with field location and wavelength, making it a proper choice for optimized on-axis and off-axis data-reduction across the spectrum. Here, we present the methodology and validate PRIME on engineering and binary Keck II telescope NIRC2 images. We also present applications of PSF model parameters retrieval using PRIME: (i) calibrate the PSF model for observations void of stars on the acquired images, i.e. optimize the PSF reconstruction process, (ii) update the AO error breakdown mutually constrained by the telemetry and the images in order to speculate on the origin of the missing error terms and evaluate their magnitude, and (iii) measure photometry and astrometry with an application to the triple system Gl569 images.

Abstract
The Apriori algorithm is considered a classic in the association rules extraction field. This algorithm makes recursive searches in a dataset looking for frequent sets that satisfy given minimum support. Apriori has several properties to optimize its performance, such as reducing the number of generated itemsets and its parallelization by multiple processors. These features have led to the emergence of several studies that present parallel versions of Apriori. However, these proposals do not explore the heterogeneous capabilities of each machine, which causes a significant part of the algorithm's processing time to be spent on I/O processes and not exactly on the execution of the algorithm. In this sense, this study proposes a mathematical modeling of the Apriori algorithm in which heterogeneous machines are considered. The findings identified a better performance of this algorithm when compared to the original and parallel versions of Apriori, but in which all processors are considered homogeneous. The findings reveal the time reducing rate increases with the growth in the number of itemsets and the number of considered processors.

Abstract
Human integration in cyber-physical systems (CPS) is playing a crucial role in the era of the digital transformation, notably because humans are seen as the most flexible driver in an automated system. Two main reference models for human activities in production systems are usually considered, namely Human-in-the-Loop (HitL) and Human-in-the-Mesh (HitM), which present different requirements and challenges. This paper aims to overview the different activities related to the human integration in CPS, particularly discussing the requirements that can be found in HitL and HitM models for the different phases of the decision-making process, namely detect, determine, develop and describe; and analyzing the technologies and computational tools to support these human activities. The human integration in CPS is illustrated through three examples, where humans playing the operator and manager roles are integrated in the PERFoRM and FAR-EDGE ecosystems, covering different phases of the decision-making process. Copyright