When an Android mobile device starts for the first time, the Android Setup Wizard (ASW) is loaded so the user can do his initial configurations, such as Wi-Fi network, date, user account, data restore, terms and condition, and others. However, in the context of manual or automated tests, every time a device starts for the first time, this ASW should be consistently configured before the tests begin. This configuration should not affect the test measurements to make them reproducible. It is impracticable to manually pass through all the steps of the ASW before every new test since this limits scalability. Moreover, a manual process can be error-prone, since two testers might use different navigation paths in the ASW, and this can affect the output. An automated script to do this task is not trivial as steps of the ASW can change depending on the device model, Android version (N, O, P, Q, etc.) and vendor (Verizon, Global, etc.). This means that the number of steps and the actions that should be executed in each of them can vary. Therefore, we propose WIZARD, an intelligent system that navigates automatically and consistently through the steps of the ASW, allowing the specialist to customize this navigation. To do that, WIZARD uses information retrieval (IR), vector space model (VSM), and re-rank techniques. We validate our approach on a set of real-world and commercially-available Android builds of a large company that comprises 1505 unique records. In the task of predicting the action to execute, our approach achieved an average Accuracy@1 of 98,82% and an Accuracy@5 of 99,87%. To the best of our knowledge, WIZARD is the first system that automatically navigates in the ASW to prepare devices for reproducible tests.
In the frame of POR FESR 2014-2020 Programs, the SoundLight project is intended to be a support for the inpatients, to ameliorate their pre and post-surgery psychophysiological conditions. SoundLight is an adaptive system, based on an intelligent agent designed for the control and the combined administration of auditory and visual stimuli, in a clinical environment. More precisely, the agent autonomously processes the sound signal, coming from various sources, and provides real-time conversion of the signal into controlled lighting effects, that are perceived relaxing by patients. The system is based on spectral signal analysis for the characterization of the sound signal and leverages on a novel sound-to-color conversion algorithm, which allows to transform main sound characters first into a chromatic reference wavelength and finally into a set of RGB values to be sent to light sources. The paper briefly introduces some essential concepts of the spectral analysis developed and focuses on the hardware/software architecture adopted to achieve real-time processing while maintaining a very compact form-factor of the processing hardware. Some key concepts related to the ongoing validation are also presented1.
Online health forums are increasingly used by patients to share information and discuss a broad range of health conditions. In this study, we investigate the presence of comorbidity of mental and physical disorders in users of such forums. We apply natural language processing methods to identify posts where users discuss mental health problems alongside chronic physical diseases, and use data mining techniques to explore the comorbidity patterns in these posts. We compare our findings to those reported in the literature and show how the results obtained are correlated with real-life events.
The current energy transition implies fundamental changes in energy systems, in both their structure and operation. In this context, microgrids have the potential to reshape the traditional power grid, empowering the role of small-scale renewable power generators towards more sustainable, cost-effective, and reliable energy systems. Microgrids are operated by energy management systems whose development is driven mainly by the contributions of technologies for distributed energy resources and techniques for optimal planning of loads. Yet, little attention is given to the design of such systems as software tools, whose capabilities are often limited because of a monolithic structure. We postulate that such systems need to intertwine both perspectives if microgrids are going to become a realised potential. We therefore propose a novel design of an energy management system that considers both perspectives and is based on the principles of service-orientation. We implement the proposed design into a service-oriented energy management system prototype and we run simulations with real data to test its feasibility.
The major part of the success that we have had in deep learning is attributed to the proper design of the architecture of the network model. With all the hardware resources and mathematical foundation at hand, it is the clever assembling of the pieces that define the performance of that deep learning model. Almost all of the research in this field till date use the models that are designed by a human researcher using their past experiences or applying the iterative process of adding components to the model to see which one performs the best. As we strive towards the general Artificial Intelligence, this process of manually tuning the network to make it work on a specific task does not add up well for the cause. We are concerned with automating the task of network architecture design where the learning algorithm tries to discover the best performing model on itself. To do this, we employ an evolving algorithm in addition to the gradient descent to evolve and train the model. Specifically, we use a form of genetic algorithm with mutation and recombination operators to constantly change the architecture, while the commonly used gradient descent is used as a learning algorithm for each of the genetically procreated models. We propose a series of mutation operators and a method of recombination called Highest Varying k-Features Recombination(HVk-FR) to evolve the CNN models. Results show using our method of recombination on top of mutation yields the best accuracy.
Almost every major city in the world is facing a significant economic loss caused by traffic congestions. In this context, it already has been shown that Adaptive Traffic Light Control (ATLC) can be an effective solution to improve a diversity of different traffic-related metrics. The problem of ATLC can be modeled in various ways with reinforcement learning being one of the most promising frameworks. Especially the application of Multi-Agent Reinforcement Learning (MARL) can be a suitable approach for learning to adaptively control the traffic of realistic road networks. Among the set of MARL algorithms, Multi-Agent Reinforcement Learning for Integrated Network (MARLIN) stands out and is shown to be particularly suited to the problem of ATLC with producing remarkable results. MARLIN models the multi-agent framework as a stochastic game providing an explicit coordination mechanism for the agents. However, in MARLIN, the possible size of the state and action space is limited and the features for the state representation need to be hand-crafted. Therefore, in this study, the algorithm is combined with function approximation by using artificial neural networks to overcome these limitations. deep-MARLIN is explained and bench-marked in a large and realistic traffic environment using Simulation of Urban MObility (SUMO). The results indicate that when compared to MARLIN, deep-MARLIN converges faster to a policy that is producing lower average vehicle delay.
Early detection of crop disease is an essential step in food security. Usually, the detection becomes possible in a stage where disease symptoms are already visible on the aerial part of the plant. However, once the disease has manifested in different parts of the plant, little can be done to salvage the situation. Here, we suggest that the use of visible and near infrared spectral information facilitates disease detection in cassava crops before symptoms can be seen by the human eye. To test this hypothesis, we grow cassava plants in a screen house where they are inoculated with disease viruses. We monitor the plants over time collecting both spectra and plant tissue for wet chemistry analysis. Our results demonstrate that suitably trained classifiers are indeed able to detect cassava diseases. Specifically, we consider Generalized Matrix Relevance Learning Vector Quantization (GMLVQ) applied to original spectra and, alternatively, in combination with dimension reduction by Principal Component Analysis (PCA). We show that successful detection is possible shortly after the infection can be confirmed by wet lab chemistry, several weeks before symptoms manifest on the plants.
Dynamic time warping (DTW) measures the dissimilarity of time series and is a fundamental technique in pattern recognition. In practice, two time series may be subject to some global geometric (e.g. affine) transformation. Thus, the DTW computation has to be extended to optimizing warping and transformation parameters simultaneously. Due to its high complexity this optimization problem is typically solved by using the Expectation Maximization (EM) algorithm. A big disadvantage of this approach, however, is that one does not explore the complete space of transformation parameters, resulting in the danger of missing the genuine optimum. In this work we solve the original optimization problem directly. A parameter decomposition technique is applied to reduce the number of optimization parameters to that of (affine) transformation parameters only, which is a fraction of all optimization parameters. Experimental results on signature data are shown to demonstrate the ability of our approach to obtain improved precision of affine DTW computation and enhanced classification performance.
The ability of a patient to take correct medicine at right time may be reduced when having visual or auditory impairments. Use of inappropriate drug intake can be dangerous and it is important that the patient takes right drug at schedule time. But it is difficult for the elderly persons and the patients with audio and visual impairments to carry out treatment process independently and correctly. This article presents a Convolutional Neural Network (CNN) based medication monitoring system and this system is a sub component of an intelligent pill reminder system.The goal of the intelligent pill reminder system in general, is to assist patient during treatment process at home and role of the monitoring system in particular, is to minimize medication errors. This system is demonstrated on GUI application with a satisfactory accuracy.
Bayesian optimization seeks the global optimum of a black-box, objective function f (x), in the fewest possible iterations. Recent work applied knowledge of the true value of the optimum to the Gaussian Process probabilistic model typically used in Bayesian optimization. This, together with a new acquisition function called Confidence Bound Minimization, resulted in a Gaussian probabilistic posterior in which the predictions were no greater than the known maximum (and no less than for minimum). Our novel work applies Confidence Bound Minimization to Bayesian optimization with Student's-t Processes, a probabilistic alternative which addresses known weaknesses in Gaussian Processes - outliers' probability and the calculation of posterior covariance. The new model is applied to the problem of hyperparameter tuning for an XGBoost classifier. Experiments show superior regret minimization and predictive accuracy, versus the popular Expected Improvement acquisition function. Combining Confidence Bound Minimization with a transformed Student's-t Process probabilistic model and known optima produces superior training regret minimization and posterior predictions for the Six-Hump Camel(2D) and Levy(4D) benchmark problems, which do not fall below true minima.
Visually impaired people face several problems in their daily life. One of the biggest problems is to visiting unfamiliar places and identifying public places like pharmacy store, restrooms, pedestrian signs on roads, etc. Although there are some conventional methods that are available to aid visually impaired people but these are inefficient to use without assistance. The proposed method presents a framework which will help visually impaired people to identify the common public amenities while visiting any unfamiliar places. This method uses deep learning for recognizing some daily used places. For this purpose, VGG16 model is used to extract features from the images and train the sequential model. The model has been tested on varying images of different class that are present in the database. The developed algorithm achieves an accuracy of 95.88%. The obtained result of the developed model shows that it is an efficient method for assisting visually impaired people in real time application.
Using reinforcement learning to find new control strategies for manufacturing processes is a promising approach. However, in order to use reinforcement learning in a manufacturing environment, industrial requirements must be met. In this paper, a new control architecture is proposed that allows reinforcement learning frameworks to be executed on programmable logic controllers, which implies requirements regarding real-time execution. An agent exchange module is presented, that allows to automatically load agents from a non real-time learning environment to a control program running in a real-time environment of a programmable logic controller. The proposed architecture is evaluated on an experimental setup with a non-linear process and commercial off-the-shelf control hardware, to meet industrial requirements. A genetic algorithm is used to learn new control strategies without having prior knowledge of the system.
In this work we show our developed model for a self-driving car that solves a road segmentation task and at the same time classifies whether the vehicle is moving in or off the lane. We demonstrate how scene classifier can be efficiently embedded into a segmentation neural network model by the means of branching additional layers with small number of trainable parameters. The model was trained and tested on our own dataset with good accuracy performance. Experiment with external dataset proves efficiency of the proposed approach and shows reasonable results in both segmenation and classification tasks.
In recent years, deep neural networks have known a wide success in various application domains. However, they require important computational and memory resources, which severely hinders their deployment, notably on mobile devices or for real-time applications. Neural networks usually involve a large number of parameters, which correspond to the weights of the network. Such parameters, obtained with the help of a training process, are determinant for the performance of the network. However, they are also highly redundant. The pruning methods notably attempt to reduce the size of the parameter set, by identifying and removing the irrelevant weights. In this paper, we examine the impact of the training strategy on the pruning efficiency. Two training modalities are considered and compared: (1) fine-tuned and (2) from scratch. The experimental results obtained on four datasets (CIFAR10, CIFAR100, SVHN and Caltech101) and for two different CNNs (VGG16 and MobileNet) demonstrate that a network that has been pre-trained on a large corpus (e.g. ImageNet) and then fine-tuned on a particular dataset can be pruned much more efficiently (up to 80% of parameter reduction) than the same network trained from scratch.
The commonplace generation of the digital patterns by the classic watermarking approach has some visible constraints. This paper proposes a significant way to generate a watermarked image using a multimodal biometric system in which a fingerprint and a palm-print biometric of a user has chosen. The generation of a digitally watermarked image is done by using features of both the biometric which has been logically combined to satisfy ownership and is robust indeed. This paper efficaciously resolves the ownership conflict of the host data. The fingerprint biometric image of a user is taken as the host data and its digital watermarked image is fabricated by embedding the palm-print of the same user, which ensures the user's ownership stamp by satisfying the dual authentication. Created watermarked image has been scrutinized through a study for its uniqueness and recognition accuracy. The approach of the proposed algorithm is to make the watermarked data reliably invulnerable while dispersing it in the shared database. Singular value decomposition (SVD) algorithm has been practiced to inculcate the palm-print watermark in the fingerprint image. Outcomes of experimentations evince that the watermarked data is qualified to endure the triggered image processing attack and keep the perceptual properties of data unvaried. The achievement of this work is its rate of perceptual transparency and robustness of the watermarked image.
In recent years, CNNs are capturing the attention of a large community of researchers, attracted by the high performance of this approach and by the surprising results obtained in many recognition/classification activities. Unfortunately, the excellent performance of CNN-based systems is accompanied by a worryingly poor understanding of why they work so well. In this document, the basic mechanisms related to the extraction of points of interest (following the first convolution phases) are considered and compared to human fixations, with the aim of better understanding analogies and differences between computational models and human recognition. Alongside, the points of interest extracted from different layers are compared in order to evaluate similarity between different networks and the how interest points evolve within the same network. Human fixations and points of interest are initially used to construct density distribution maps; a novel similarity index is then proposed in order to compare these distribution maps. Experiments on the ETD database show that human fixations, on average, tend to be contained in the density of CNN points. In other words, human fixations seem to somehow optimize the active exploration of the image by properly using the information coming from the periphery of the visual field. These first interesting results could condition emerging models for visual recognition and visual indexing, pushing towards a more attentive exploitation of coarse structural information.
For forecasting of energy related time series data (e.g. "load" or "generation"), many different kind of learning algorithms exist. The task of selecting an appropriate algorithm which is adequate for usage with a pre-given time series is not easy for non-data science experts. Since a trial-and-error approach for finding a suitable algorithm is tedious, computationally intensive and time-consuming, meta learning approach which describes how to design an appropriate methodology to constrain the search space for fully automatically finding the suitable learning algorithm is proposed. In the present paper, based on the assumption that good indicators describing certain characteristics of an e.g. energy time series dataset may provide useful insight into which forecasting algorithms are most suitable, a Descriptive Statistics Time-based Meta Features (DSTMF) description format for energy time series datasets is proposed to efficiently select an appropriate learning algorithm to perform energy forecasting on a time series dataset. In order to demonstrate the performance of the new methodology, experiments on datasets which are load time series from 60 institutional buildings are conducted. Based on a similarity-based clustering analysis, the potential of DSTMF's meta features for capturing deep characteristics of energy time series datasets is evaluated and compared to other state-of-the-art meta features. The experiments show very good results and outperform state-of-the-art meta features to enhance model selection in energy time series forecasting.
Tissue converting lines represent one of the key plant in the paper production field: with them, paper tissue is converted into its final form for domestic and sanitary usage. One of the key points of the tissue converting lines is the productivity and the possibility to follow conversion process at relativity low cost. Despite the actual lines have yet an high productivity, the study of the state of the art has shown that choke points still exist, caused by inadequate automation. In this paper, we present the preliminary results of a project which aims at removing such obstacle towards complete automation, by introducing a set of innovations based on ICT solutions applied to advanced automation. In detail, advanced computer vision and video analytics methods will be applied to pervasively monitor converting lines and to automatically extract process information in order to self-regulate specific machine and global parameters. Big data analysis methodologies will be also integrated to obtain new knowledge and infer optimal management models which could be used for the predictive maintenance. Augmented reality interfaces are being designed and developed to support converting line monitoring and maintenance, both ordinary and extraordinary. An Artificial Intelligence module provides suggestions and instructions to the operators in order to guarantee production level even in case of unskilled staff. The automation of such processes will improve factory safety, decrease manual interventions and, thus, will increase production line up-time and efficiency.
This paper is of a conceptual nature and focuses on the use of a specific virtual reality environment in civil-military training. We analyzed the didactic potential of so-called CAVE automatic virtual environments for First Responder training, a type of training that fills the gap between First Aid training and the training received by emergency medical technicians. Since real training involves live drills based on unexpected situations, it is expensive and difficult to organize. We propose the application of virtual environments of three different sizes to increase the effectiveness of First Responder training. Our findings show that using the CAVEs allows for easier adaptation of the learning environment to a specific scenario while reducing costs from preparing training fields. The environments are also safer for trainees who must perform tasks related to hazardous materials. The analysis was based on the facilities at the Immersive 3D Visualization Lab (I3DVL) at Gdańsk University of Technology in Poland. We believe that, in the context of emerging threats of chemical, biological, radiological, and nuclear terrorism, such training is a necessity. The application of artificial intelligence can be considered as a further step in the facility development to increase the training fidelity.
GEMTEX Lab has made some efforts in the application of artificial intelligence to optimize the garment manufacturing process (Zeng, Ge & Bruniaux, 2008; Thomassey & Zeng, 2018), including cutting, sewing, ironing and packing. As for the cutting process, marker making plays an essential role. The exponentially increasing magnitude of possible size combinations due to a considerable larger size number in modern garment mass customization induce a larger workload of marker making. Compared with creating all the markers using commercial marker making software, the application of machine learning technologies for marker length estimation will benefit in both efficiency and accuracy. The results generated from the marker length estimation can be used for providing the input data of cutting order planning (COP) for lean garment production, and as well predicting the objective values of marker length for the guidance and evaluation of marker making. In this study, MLR and RBF NN were used to estimating marker lengths with various sizes (regarding mass production (MP) and mass customization (MC)). The results showed that MLR generally outperforms RBF NN. MLR is more appropriated for general markers containing articles of regular sizes and mixed markers. RBF NN can be potential for markers containing articles of irregular sizes or for group markers.
