Saberian M.,Netflix |
Pereira J.C.,INESCTEC |
Xu C.,Google |
Yang J.,Yahoo! |
Vasconcelos N.,University of California at San Diego
Advances in Neural Information Processing Systems | Year: 2016
In this paper we establish a duality between boosting and SVM, and use this to derive a novel discriminant dimensionality reduction algorithm. In particular, using the multiclass formulation of boosting and SVM we note that both use a combination of mapping and linear classification to maximize the multiclass margin. In SVM this is implemented using a pre-defined mapping (induced by the kernel) and optimizing the linear classifiers. In boosting the linear classifiers are pre-defined and the mapping (predictor) is learned through a combination of weak learners. We argue that the intermediate mapping, i.e. boosting predictor, is preserving the discriminant aspects of the data and that by controlling the dimension of this mapping it is possible to obtain discriminant low dimensional representations for the data. We use the aforementioned duality and propose a new method, Large Margin Discriminant Dimensionality Reduction (LADDER) that jointly learns the mapping and the linear classifiers in an efficient manner. This leads to a data-driven mapping which can embed data into any number of dimensions. Experimental results show that this embedding can significantly improve performance on tasks such as hashing and image/scene classification. © 2016 NIPS Foundation - All Rights Reserved.
News Article | December 21, 2016
Renewable energy was responsible for producing this year around 59% of the total amount of electricity in Portugal and it is expected that 1/3 comes from wind energy sources, according to data disclosed by APREN. It becomes more and more crucial to support alternative solutions to produce energy, a topic that was highlighted on the 2016 Portuguese Association of Renewable Energies Award which has honoured Bernardo Marques Amaral Silva, a INESCTEC researcher and a MIT Portugal Program Sustainable Energy Systems Phd doctorate at the Faculty of Engineering of the University of Porto with the first prize for his thesis "Multi-terminal HVDC Grids: Control Strategies for Ancillary Services Provision in Interconnected Transmission Systems with Offshore Wind Farms" , a pioneering approach to take the best profit of Wind farms (WF). Wind Energy (WE) has largely contributed to the de-carbonisation of the energy sector and consequently to the de?nition of the European Commission (EC) targets on renewable-based electricity generation. During the last decade, massive investment has culminated in the substantial installation of Wind Farms (WF). Moreover, the ambitious plans for increasing these targets on renewable-based electricity generation demand the deployment of more WF. However, the European Union has settled the goal for the level of greenhouse gas reduction by 40% (by comparison with 1990 levels) until 2030 and the share of renewable energy to 27% in terms of energy consumption. Simultaneously, the Fukushima incident has put pressure on European governments regarding the use of nuclear power plants which, although not considered as renewable, have reduced emission of greenhouse gases (compared to conventional power plants based on fossil fuels). In order to meet the new goals, it is expected a significant contribution from offshore wind farms (at sea). From a technical point of view, the adoption of High Voltage Direct Current (HVDC) is crucial to allow the installation of high power PE and to manage high connexion distances from the mainland. Recent procedures intend to adopt direct current networks rather than connections "point-to-point" since they may also support the connection between mainland AC networks, allowing a greater integration of renewable sources and enabling the development of a European electricity market. The work developed by Bernardo Marques Amaral Silva has focused on the development of strategies to control and operate HVDC networks with offshore wind farms, to allow the delivery of support services to the mainland electrical systems, in an autonomous environment without communication system. In a first phase, a control scheme for allowing offshore WF as well as the interconnected AC mainland grids on participating on primary frequency control has been developed. This control scheme also took into account the need of provision of synthetic inertia by offshore WF. The results showed a successful accomplishment on mitigating an interconnected mainland grid frequency disturbance. In a second phase, DC grids control schemes were developed and tested aiming at the operation of the DC grid during an AC mainland fault event (Fault Ride-Through - FRT capability) and also to assure the operation of the DC grid following a permanent fault events or the disconnection of an onshore converter. According to the former MIT Portugal Program student, the major outcome of this project consists on allowing the management of more wind energy production and consequently increasing energy coming from renewable sources. The idea has already been tested in laboratory and its feasibility has been verified so, in a near future, the goal is to test it in an existing wind farm. The main objective of the second edition of the APREN Award was to support the best and most relevant academic Thesis held in Portuguese higher education institutions (related to renewable electricity) and to contribute to the disclosure and knowledge transfer between Research Centers and companies.
De Sa C.R.,INESCTEC |
De Sa C.R.,Leiden University |
Soares C.,INESCTEC |
Knobbe A.,Leiden University |
And 2 more authors.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) | Year: 2013
Label Ranking (LR) problems, such as predicting rankings of financial analysts, are becoming increasingly important in data mining. While there has been a significant amount of work on the development of learning algorithms for LR in recent years, pre-processing methods for LR are still very scarce. However, some methods, like Naive Bayes for LR and APRIORI-LR, cannot deal with real-valued data directly. As a make-shift solution, one could consider conventional discretization methods used in classification, by simply treating each unique ranking as a separate class. In this paper, we show that such an approach has several disadvantages. As an alternative, we propose an adaptation of an existing method, MDLP, specifically for LR problems. We illustrate the advantages of the new method using synthetic data. Additionally, we present results obtained on several benchmark datasets. The results clearly indicate that the discretization is performing as expected and in some cases improves the results of the learning algorithms. © 2013 Springer-Verlag.
Baptista R.,INESCTEC |
Coelho A.,INESCTEC |
De Carvalho C.V.,ISEP
Proceedings of the European Conference on Games-based Learning | Year: 2015
Recently, Serious Games (SG) achieved a recognized position as a learning tool in several contexts. SG provide constructive learning environments in which errors can be made without real life penalties and where students get instant feedback from their actions when facing challenges. These challenges should be in accordance with the intended learning goals and they should adapt and/or be repeated according to the learner's level. This aspect is decisive in the acquisition of knowledge, experience and professional skills through the simulation of different situations and contexts. The effectiveness of competences' training is directly related to the success in their acquisition but, above all, it is related to the ability to apply them to successfully perform a given task. However, an optimal game design methodology for competence training is yet to be created. This article presents a study that identifies the most appropriate game categories to develop specific skills and competences. It considers a taxonomy with eight game categories (Action, Strategy, Playing, Sports, Management Simulation, Adventure, Puzzle and Quiz) that were matched with the Education Competences and Educational Competency Wheel. Analysing 116 serious games allowed identifying which categories were more efficient in the training of a specific competence and therefore should be reused in the same scope.
Carvalho L.,INESCTEC |
Roriz P.,INESCTEC |
Frazao O.,INESCTEC |
Frazao O.,University of Porto |
And 2 more authors.
Journal of Physics: Conference Series | Year: 2015
Cross-bite, as a malocclusion effect, is defined as a transversal changing of the upper dental arch, in relation to the lower arch, and may be classified as skeletal, dental or functional. As a consequence, the expansion of maxilla is an effective clinical treatment used to correct transversal maxillary discrepancy. The maxillary expansion is an ancient method used in orthodontics, for the correction of the maxillary athresia with posterior crossbite, through the opening of the midpalatal suture (disjunction), using orthodontic- orthopaedic devices. Same controversial discussion arises among the clinicians, about the effects of each orthodontic devices as also about the technique to be employed. The objective of this study was to compare the strain field induced by two different orthodontic devices, named disjunctor with and without a connecting bar, in an acrylic model jaw, using fiber Bragg grating sensors to measure the strain patterns. The orthodontic device disjunctor with the bar, in general, transmits higher forces and strain to teeth and maxillae, than with the disjunctor without bar. It was verified that the strain patterns were not symmetric between the left and the right sides as also between the posterior and anterior regions of the maxillae. For the two devices is also found that in addition a displacement in the horizontal plane, particularly in posterior teeth, also occurs a rotation corresponding to a vestibularization of the posterior teeth and their alveolar processes. © Published under licence by IOP Publishing Ltd.
Costa C.M.,INESCTEC |
Sobreira H.M.,INESCTEC |
Sousa A.J.,INESCTEC |
Robotics and Autonomous Systems | Year: 2016
Mobile robot platforms capable of operating safely and accurately in dynamic environments can have a multitude of applications, ranging from simple delivery tasks to advanced assembly operations. These abilities rely heavily on a robust navigation stack, which requires stable and accurate pose estimations within the environment. To solve this problem, a modular localization system suitable for a wide range of mobile robot platforms was developed. By using LIDAR/RGB-D data, the proposed system is capable of achieving 1-2 cm in translation error and 1°-3° degrees in rotation error while requiring only 5-35 ms of processing time (in 3 and 6 DoF respectively). The system was tested in three robot platforms and in several environments with different sensor configurations. It demonstrated high accuracy while performing pose tracking with point cloud registration algorithms and high reliability when estimating the initial pose using feature matching techniques. The system can also build a map of the environment with surface reconstruction and incrementally update it with either the full field of view of the sensor data or only the unknown sections, which allows to reduce the mapping processing time and also gives the possibility to update a CAD model of the environment without degrading the detail of known static areas due to sensor noise. © 2015 Elsevier B.V. All rights reserved.