Designing a Convolutional Neural Networks (CNN) is a complex task and requires expert knowledge to optimize the performance and network architecture. In this paper, a novel data-driven approach is proposed to determine the architecture of CNN models. The proposed Autonomous Convolutional Neural Networks (AutoCNNThe executable code and original numerical results can be downloaded from (https://tinyurl.com/AutoCNN)) algorithm introduces data driven strategies for addition of new convolutional layers, pruning of redundant filters and training cycle optimization. AutoCNN is evaluated using MNIST, MNIST-rot-back-image, Fashion MNIST and the ADHD200 datasets to measure the performance on small datasets with varied feature distributions. The results indicate that AutoCNN optimizes the CNN network architecture and helps maximise the classification performance. The data-driven network determination approach introduced in this paper was found to not only provides competitive performance similar to existing evolutionary computation based network determination algorithms in literature, but was found to be an effective optimization tool to improve the performance of existing CNN architectures. Further, the AutoCNN was found to highly immune to noise in the dataset and has proven to be effective method to transfer knowledge between related datasets. Therefore, the AutoCNN is a highly versatile CNN architecture determination tool that has a wide range of applications in the field of autonomous driving, medical image analysis, image enhancement, camera based security monitoring and image based fault detection

The Intergovernmental Panel on Climate Change and the Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services have emphasised unequivocal evidences about the impact of human actions on climate and biodiversity at alarming rates. In Brazilian terms, 2019 has been marked by controversial discussions among politicians and environmentalists, leading to misinformation and misinterpretations that clearly motivate the continuous collection and scientific analysis of data to support sustainable solutions. Aiming at dealing with this issue, this manuscript brings two contributions: (i) the creation of the Brazilian Forest Dataset, including Brazilian seed plants, Fraction of Absorbed Photosynthetically Active Radiation, meteorological and geographical data composing 8,482 attributes to model and predict 20 vegetation types; and (ii) the feasibility analysis on modelling this dataset in light of supervised machine learning algorithms, so we devise confident results on the Brazilian biodiversity. Experimental results confirm Random Forest and Support Vector Machines successfully adjust models, enabling researchers to predict the occurrence of specific types of vegetation in different regions of Brazil as well as analyse how the prediction accuracy changes along time after the collection of new data. Our contributions bring important tools to support the study on the evolution of the Brazilian biodiversity.

Most of the existing outsourced encrypted data schemes are retrieved based on the query keyword entered by authorised users.However, with the increase of the data scale in the cloud storage system, the retrieval efficiency of existing solutions has not been significantly improved.In this paper, a multi-keyword ranked search scheme for ciphertext based on mapping set matching (MSMR) is proposed, where (1) The private cloud server matches the keyword numbering set corresponding to the document index vector and the keyword numbering set corresponding to the query vector and sends the document identifier of the matching keyword numbering to the public cloud server.The public cloud server filters the documents irrelevant to the query request according to the document identifier corresponding to the matching keyword numbering, which effectively reduces the time spent in calculating the correlation score, and (2) the document index vector and query vector are segmented before encrypting them out, reducing the time to construct such vectors.Theoretical analysis shows that the proposed scheme is secure in the known ciphertext model.Experimental results confirm that whenever the data scale grows, the improvement of MSMR retrieval efficiency is more significant.

We demonstrate that projecting data points into hyperplanes is good strategy for general-purpose kernel design. We used three different hyperplanes generation schemes, random, convex hull and α-shape, and evaluated the results on two synthetic and three well known image-based datasets. The results showed considerable improvement in the classification performance in almost all scenarios, corroborating the claim that such an approach can be used as a general-purpose kernel transformation. Also, we discuss some connection with Convolutional Neural Networks and how such an approach could be used to understand such networks better.

International audience

Assigning scores to individual features is a popular method for estimating the relevance of features in supervised learning. An accurate feature score estimation provides essential insights in sensitive domains, which is decisive to explain how features influence a given decision, contributing to the interpretability of the model. Learning from streaming data adds several challenges to machine learning tasks, including limited resources and changes to the underlying data distribution (i.e., evolving data streams). In this work, we introduce and analyze methods to efficiently estimate the Mean Decrease in Impurity (MDI) and COVER measures using ensembles of incremental decision trees. To achieve current scores in evolving data streams, we employ tree-ensembles that incorporate active drift detection. Experimental results show how MDI and COVER can be used to track the feature scores when their importance to the ensemble model shift over time. On top of that, we present the impact on the feature scores when the learning problem includes a non-negligible verification latency for the arrival of the labels. We also present a counter-intuitive experiment using a standard benchmark dataset where the feature scores correctly illustrate the importance of two features to the ensemble model. However, these features are prioritized due to biased split decisions, and in their absence, the model increases in predictive performance. We conclude that the presented measures can be used to understand the impact of features in the ensemble model better, still, such measures should be used with caution as they are limited by the underlying tree building and ensemble model biases.

International audience

International audience

The Convolutional Neural Network (CNN) figures among the state-of-the-art Deep Learning (DL) algorithms due to its robustness to support data shift, scale variations, and its capability of extracting relevant information from large-scale input data. However, setting appropriate parameters to define CNN architectures is still a challenging issue, mainly to tackle real-world problems. A typical approach consists in empirically assessing different CNN settings in order to select the most appropriate one. This procedure has clear limitations, including the choice of suitable predefined configurations as well as the high computational cost involved in evaluating each of them. This work presents a novel methodology to tackle the previously mentioned issues, providing mechanisms to estimate effective CNN configurations, including the size of convolutional masks (convolutional kernels) and the number of convolutional units (CNN neurons) per layer. Based on the False Nearest Neighbors (FNN), a well-known tool from the area of Dynamical Systems, the proposed method helps estimating CNN architectures that are less complex and produce good results. Our experiments confirm that architectures estimated through the proposed approach are as effective as the complex ones defined by empirical and computationally intensive strategies.

The Statistical Learning Theory (SLT) defines five assumptions to ensure learning for supervised algorithms. Data independency is one of those assumptions, once the SLT relies on the Law of Large Numbers to ensure learning bounds. As a consequence, this assumption imposes a strong limitation to guarantee learning on time-dependent scenarios. In order to tackle this issue, some researchers relax this assumption with the detriment of invalidating all theoretical results provided by the SLT. In this paper we apply a kernel function, more precisely the Takens' immersion theorem, to reconstruct time-dependent open-ended sequences of observations, also referred to as data streams in the context of Machine Learning, into multidimensional spaces (a.k.a. phase spaces) in attempt to hold the data independency assumption. At first, we study the best immersion parameterization for our kernel function using the Distance-Weighted Nearest Neighbors (DWNN). Next, we use this best immersion to recursively forecast next observations based on the prediction horizon, estimated using the Lyapunov exponent. Afterwards, predicted observations are compared against the expected ones using the Mean Distance from the Diagonal Line (MDDL). Theoretical and experimental results based on a cross-validation strategy provide stronger evidences of generalization, what allows us to conclude that one can learn from time-dependent data after using our approach. This opens up a very important possibility for ensuring supervised learning when it comes to time-dependent data, being useful to tackle applications such as in the climate, animal tracking, biology and other domains.

High-dimensional data streams clustering is an attractive research topic, as there are several applications that generate a high number of attributes, bringing new challenges in terms of partitioning due to the curse of dimensionality. In addition, those applications produce unbounded sequences of data which cannot be stored for later analysis. Although the importance of this scenario, there are still very few algorithms available in the literature to meet this task. Despite the theoretical foundation of mathematical topology for dealing with high-dimensional spaces, none of those approaches have investigated the problem of finding topologically similar projected clusters in high-dimensional data streams. Among the advantages of topology is the possibility to analyze data in a coordinate-free and noise-robust manner. In a previous research, we have shown that topologically similar clusters can be meaningful considering real-world data sets. In this paper, we extend those ideas and propose PTS, an algorithm for finding topologically similar clusters in high-dimensional data streams. The algorithm is capable of finding traditional projected clusters and then merging them according to topological features computed using persistent homology. Experiments with synthetic data streams of dimensions d=8,16,32,64 and 128 confirm the ability of PTS to find topologically similar projected clusters.

The ability to detect changes in the data distribution is an important issue in Data Stream mining. Detecting changes in data distribution allows the adaptation of a previously learned model to accommodate the most recent data and, therefore, improve

Several research fields have described phenomena that produce endless sequences of samples, referred to as data streams. These phenomena are studied using data clustering models continuously obtained throughout the endless data gathering process, who

Current scientific applications have been producing large amounts of data. The processing, handling and analysis of such data require large-scale computing infrastructures such as clusters and grids. In this area, studies aim at improving the performance of data-intensive applications by optimizing data accesses. In order to achieve this goal, distributed storage systems have been considering techniques of data replication, migration, distribution, and access parallelism. However, the main drawback of those studies is that they do not take into account application behavior to perform data access optimization. This limitation motivated this paper which applies strategies to support the online prediction of application behavior in order to optimize data access operations on distributed systems, without requiring any information on past executions. In order to accomplish such a goal, this approach organizes application behaviors as time series and, then, analyzes and classifies those series according to their properties. By knowing properties, the approach selects modeling techniques to represent series and perform predictions, which are, later on, used to optimize data access operations. This new approach was implemented and evaluated using the OptorSim simulator, sponsored by the LHC-CERN project and widely employed by the scientific community. Experiments confirm this new approach reduces application execution time in about 50 percent, specially when handling large amounts of data.

Real-world datasets commonly present high dimensional data, which means an increased amount of information. However, this does not always imply an improvement in learning technique performance. Furthermore, some features may be correlated or add unexpected noise, thereby reducing data clustering performance. This has motivated the development of feature selection methods to find the most relevant subset of features to describe data. In this work, we focus on the problem of unsupervised feature selection. The main goal is to define a method to identify the number of features to select after sorting them based on some criterion. This task is done by means of the False Nearest Neighbor technique, which is rooted in chaos theory. Results have shown that this technique gives a good approximate number of features to select. When compared to other techniques, in most of the analyzed cases, it maintains the quality of the generated partitions while selecting fewer features.

In this paper we propose the use of two approaches to tackle the detection and comprehension of how a faulty environment affects process behavior. We model processes as state-transition models with associated transition probabilities. The study employs SWIFI (Software Implemented Fault Injection) in commonly used UNIX tools to analyze behavioral changes. The approaches complement each other as one allows the detection of dissimilar behaviors and the other the analysis of how relationships among process states evolve over time. Obtained results are promising, as they allow a clear characterization of process execution in a normal and a faulty environment.

In order to take advantage and profit with the popularization of digital music, companies started marketing licensed content on high-storage portable media players. The introduction of wireless technology in such players motivates new business opportunities where music distribution is ubiquitous. However, in such high-supply scenario, consumers may have difficulties to find interesting content. In such context, music recommender systems assist consumers in identifying their preferences and in supporting content searches. An important feature in such market is the low attention given to new music styles, what increases the promotion costs. In order to assist consumers who, positive or negatively, pay attention to such novelty factor, this work proposes a novel method to estimate music preference profiles based on acoustic similarity measures. Such profiles are learnt by an artificial neural network, named self-organizing novelty detection neural network architecture (SONDE), which classifies and quantifies the novelty level of music titles regarding the user profile. Based on novelty levels, we suggest a discount rate model to support promotion strategies. The proposed method is evaluated by simulating some scenarios.

A grid scheduling algorithm inspired on the simulated annealing meta-heuristic (SA) is presented. Guided by predictions about parallel applications resource usage, the algorithm uses SA to find a near-optimal solution which minimizes the overall execution time for scheduling problems on heterogeneous grids. The scheduling algorithm invalidated by simulations, using a model which considers the mainly details about the distributed computers and the jobs,and is compared with other scheduling algorithms. The results allows identify the best SA parameters values, that are the learning rate and the iteration size, for each grid size. We found good results when simulating the scheduling algorithm for grid sizes near by 512 computers. We also observed that the SA computing cost do not jeopardize the scheduling results, which are better when compared with the performance of other scheduling disciplines.

Motivated by the first process allocation limitations of the original Route load balancing algorithm, this paper presents RouteGA (Route with Genetic Algorithm support) which considers historical information about parallel application executions in order to optimize the first scheduling. This information is extracted by using monitors and summarized in a knowledge base used to quantify process occupation at the launch moment. Such occupation is used to parameterize a genetic algorithm responsible for optimizing the process allocation on heterogeneous computing environments such as Grids. Results confirm RouteGA over- performs the original Route, which had previously overper-formed others from literature.

The performance of distributed computing is usually limited by the heterogeneous nature of distributed systems, data transfer rate and access latency. Techniques such as adaptive process migration, data caching and prefetching were developed to overcome this limitation. However, such techniques require the knowledge of application behavior in order to be effective. In this sense, we intend to propose a new model for application behavior prediction that, by classifying and analyzing application access patterns, is able to predict future application behavior. The model aims to allow a transparent and automatic process behavior extraction, classification and prediction, using a variable set of techniques, including stochastic models and artificial intelligence-based approaches.

The Internet evolution has been requiring the development of new technology to support multimedia transmission such as images, database access, audio and video in real-time. Such development needs new services and supports like the voice over IP (Voip) which has a main motivation in the low cost communication and management. Voip systems have motivated this work which proposes an adaptive audio quantizer named IQ (Intervalar Quantizer) to reduce the data dimensionality and consequently the entropy, what allows better audio compression. This quantizer is adaptive because it has an error tolerance parameter which can be varied according to the available network bandwidth, allowing to adapt communication. After transmitting, the audio is improved by using a filter with complex poles in the Z plan. This filter attenuates non-important frequencies, privileging the sensitive ones to human audition. Results confirm that IQ and the filter offer good quality (measured using the mean opinion score metrics) and compress ratio.

Telemedicine, distance learning and e-commerce applications impose time constraints directly related to the efficacy of their operations. In order to offer reliability levels capable of meeting such requirements, mechanisms to provide QoS have been widely employed, what motivates this work to propose, implement and validate a real-time scheduling policy for providing absolute QoS for web services. The policy, named Exigency-Based Scheduling (EBS), intends to fast serve the most urgent requests, without degrading the whole system service. The current approach is based on the real-time scheduling, low latency and feedback scheduling, allowing a balanced configuration by the quantification of the exigency imposed to the system by the service classes. The technique evaluation uses metrics proposed in the present work. Experimental results confirm improvements in terms of QoS and client satisfaction.

The evolution of computers and networks has motivated the development of high performance systems using the distributed system concepts. Such evolution has motivated the study of load balancing techniques, low latency protocols, operating systems and middlewares. Furthermore, the evolution from distributed systems based on clusters to grid environments has motivated new researches to define load balancing algorithms to support scalable and heterogeneous computing capacity environments. In this paper, the route load balancing algorithm is proposed, presented and evaluated. This algorithm is designed to equally distribute the workload of tasks of parallel applications over grid computing environments. Experiments compare the performance of the proposed algorithm to others from literature. The obtained results allow concluding that the route algorithm is indicated for environments where there are several heterogeneous computers and parallel applications are composed of multiple tasks.

This paper presents a technique to select subsets of the test cases, reducing the time consumed during the evaluation of a new software version and maintaining the ability to detect defects introduced. Our technique is based on a model to classify test case suites by using an ART-2A selforganizing neural network architecture. Each test case is summarized in a feature vector, which contains all the relevant information about the software behavior. The neural network classifies feature vectors into clusters, which are labeled according to software behavior. The source code of a new software version is analyzed to determine the most adequate clusters from which the test case subset will be selected. Experiments compared feature vectors obtained from all-uses code coverage information to a random selection approach. Results confirm the new technique has improved the precision and recall metrics adopted.

This paper presents a new model for evaluation of the positive and negative impacts related to the process migration on environments composed by heterogeneous capacity computers. On this model, a busy computer analyzes the occupation of each process and selects the more adequate for migration. The analysis and selection are done through a migration factor. This factor reflects how much the busy computer will be freed and how much the destination computer will be overloaded, in view of the migration of each process. The migrated processes are the ones that present migration factors to enhance the environment load balancing. The results from the carried out experiments have proved this model contributions when compared to related work. The contribution is the decrease in process average response time, which means higher performance.

This paper presents JCM (Java cluster management) project which provides a performance, tool-set and viability analysis of the Java platform to design high-performance applications. This is done through the design of a Java cluster management service based on the evolution of the Java platform, using MAJC (Microprocessor Architecture for Java Computing). The design of the cluster was proposed using queuing theory and UML. The JCM makes a comparison to existent clusters like Beowulf and DSP systems to the cluster management service developed.