Search

Current research in natural language processing is highly dependent on carefully produced corpora. Most existing resources focus on English; some resources focus on languages such as Chinese and French; few resources deal with more than one language. This paper presents the Pir\'a dataset, a large set of questions and answers about the ocean and the Brazilian coast both in Portuguese and English. Pir\'a is, to the best of our knowledge, the first QA dataset with supporting texts in Portuguese, and, perhaps more importantly, the first bilingual QA dataset that includes this language. The Pir\'a dataset consists of 2261 properly curated question/answer (QA) sets in both languages. The QA sets were manually created based on two corpora: abstracts related to the Brazilian coast and excerpts of United Nation reports about the ocean. The QA sets were validated in a peer-review process with the dataset contributors. We discuss some of the advantages as well as limitations of Pir\'a, as this new resource can support a set of tasks in NLP such as question-answering, information retrieval, and machine translation., Comment: https://github.com/C4AI/Pira

Computerized adaptive testing (CAT) based on item response theory allows more accurate assessments with fewer questions than the classic paper and pencil (P&P) test. Nonetheless, the CAT construction involves some key questions that, when done properly, can further improve the accuracy and efficiency in estimating the examinees’ abilities. One of the main questions is in regard to choosing the item selection rule (ISR). The classic CAT makes exclusive use of one ISR. However, these rules have differences depending on the examinees’ ability level and on the CAT stage. Thus, the objective of this work is to reduce the dichotomous test size which is inserted in a classic CAT with no significant loss of accuracy in the estimation of the examinee’s ability level. For this purpose, we analyze the ISR performance and then build a personalized item selection process in CAT considering the use of more than one rule. The case study in Mathematics and its Technologies test of the ENEM 2012 shows that the Kullback-Leibler information with a posterior distribution (KLP) has better performance in the examinees’ ability estimation when compared with Fisher information (F), Kullback-Leibler information (KL), maximum likelihood weighted information (MLWI), and maximum posterior weighted information (MPWI) rules. Previous results in the literature show that CAT using KLP was able to reduce this test size by 46.6% from the full size of 45 items with no significant loss of accuracy in estimating the examinees’ ability level. In this work, we observe that the F and the MLWI rules performed better on early CAT stages to estimate examinees’ proficiency level with extreme negative and positive values, respectively. With this information, we were able to reduce the same test by 53.3% using the personalized item selection process, called ALICAT, which includes the best rules working together.

Reinforcement Learning is a general approach to learn acting optimally in a stochastic environment; however, because of a trial-and-error strategy, learning may take prohibitive long times. An effective method to accelerate the learning process is to make use of spatial abstractions to produce a generalization of experiences. Therefore, how to automatically design abstractions is an important question in Reinforcement Learning. We formalize the problem of designing abstractions as the problem of selecting abstractions, and we propose to evaluate the learning process as a surrogate of the quality of a subset of abstractions. Because evaluating the learning process is costly, we present an algorithm, the SEPO algorithm, that construct a ranking of abstractions efficiently. To evaluate the efficiency of the SEPO algorithm, experiments are done in a soccer simulator.

The use of computer vision to analyze human movement has been growing considerably, facilitated by the increased availability of depth cameras This paper describes the results of a systematic review about the techniques used for movement tracking and recognition, focusing on metrics to compare choreographed movements using Microsoft Kinect as a sensor. Several techniques for data analysis and pattern recognition are explored for this task, particularly Dynamic Time Warping and Hidden Markov Models. Most papers we discuss used a single sensor instead of more complex setups and most took advantage of the Kinect SDK instead of alternatives. Rhythm is rarely considered in these systems due to the temporal alignment strategies used. While most systems that use the sensors for some form of interaction instead claim that this interaction is natural, very few actually perform any sort of usability or user experience analysis.

In this systematic review we approach the problem of light variation in tasks related to automatic facial recognition, a feature that can significantly affect the performance of automatic systems. We then carry out a broad research on the state of the art, describing and comparing current research in this subject. The review relies on a set of processes for searching, analysing and describing research that is considered relevant to this work, and which are reported in more detail in this article. In analysing the results, we could notice that the problem of light variation is still one of the greatest challenges in the area of facial recognition, which translates in a good deal of research directly tackling this problem, trying to solve it or, at least, mitigate it somehow, so as to improve the performance of facial recognition techniques and algorithms. Finally, results also show that this is a problem of great concern by researchers in the field, insomuch that even those articles that do not directly deal with it still make explicit the researchers' concern about it.

The ability to automatically determine a human body’s sequence of postures during movement has many practical applications, from the evaluation of the performance of physical activity practitioners to the evaluation and design of the user experience in certain systems. Current representations for such postures, however useful, are not capable of capturing all necessary features for a complete description of a human stance, such as the relationship between non-directly connected body parts. In this article, we introduce a mark-up language for body stance and movement description, designed to allow for the unambiguous representation of movement as well as the extraction of relationships between directly and non-directly connected body parts. Along with the language, we also present a computer program, developed to help end users in codifying stance and movement without having to know the language in detail.

Reinforcement Learning is carried out on-line, through trial-and-error interactions of the agent with the environment, which can be very time consuming when considering robots. In this paper we contribute a new learning algorithm, CFQLearning, which uses macro-states, a low-resolution discretisation of the state space, and a partial-policy to get around obstacles, both of them based on the complexity of the environment structure. The use of macro-states avoids convergence of algorithms, but can accelerate the learning process. In the other hand, partial-policies can guarantee that an agent fulfils its task, even through macro-state. Experiments show that the CFQ-Learning performs a good balance between policy quality and learning rate.

Ad Networks connect advertisers to websites that want to host advertisements. When users request websites, the Ad Network decides which ad to send so as to maximize the number of ads that are clicked by users. Due to the difficulty in solving such maximization problems, there are solutions in the literature that are based on linear programming (LP) relaxations. We contribute with a formulation for the Ad Network optimization problem, where it is cast as a Markov decision process (MDP). We analyze theoretically the relative performance of MDP and LP solutions. We report on an empirical evaluation of solutions for LP relaxations, in which we analyze the effect of problem size in performance loss compared to the MDP solution. We show that in some configurations, the LP relaxations incur in approximately 58 % revenue loss when compared to MDP solutions. However, such relative loss decreases as problem size increases. We also propose new heuristics to improve the use of solutions achieved by LP relaxation. We show that solutions obtained by LP relaxations are suitable for Ad Networks, as the performance loss introduced by such solutions are small in large problems observed in practice.

Reinforcement learning (RL) enables an agent to learn behavior by acquiring experience through trial-and-error interactions with a dynamic environment. However, knowledge is usually built from scratch and learning to behave may take a long time. Here, we improve the learning performance by leveraging prior knowledge; that is, the learner shows proper behavior from the beginning of a target task, using the knowledge from a set of known, previously solved, source tasks. In this paper, we argue that building stochastic abstract policies that generalize over past experiences is an effective way to provide such improvement and this generalization outperforms the current practice of using a library of policies. We achieve that contributing with a new algorithm, AbsProb-PI-multiple and a framework for transferring knowledge represented as a stochastic abstract policy in new RL tasks. Stochastic abstract policies offer an effective way to encode knowledge because the abstraction they provide not only generalizes solutions but also facilitates extracting the similarities among tasks. We perform experiments in a robotic navigation environment and analyze the agent's behavior throughout the learning process and also assess the transfer ratio for different amounts of source tasks. We compare our method with the transfer of a library of policies, and experiments show that the use of a generalized policy produces better results by more effectively guiding the agent when learning a target task.

In this paper we improve learning performance of a risk-aware robot facing navigation tasks by employing transfer learning; that is, we use information from a previously solved task to accelerate learning in a new task. To do so, we transfer risk-aware memoryless stochastic abstract policies into a new task. We show how to incorporate risk-awareness into robotic navigation tasks, in particular when tasks are modeled as stochastic shortest path problems. We then show how to use a modified policy iteration algorithm, called AbsProb-PI, to obtain risk-neutral and risk-prone memoryless stochastic abstract policies. Finally, we propose a method that combines abstract policies, and show how to use the combined policy in a new navigation task. Experiments validate our proposals and show that one can find effective abstract policies that can improve robot behavior in navigation problems.

Assessment of video segmentation quality is a problem seldom investigated by the scientific community. Nevertheless, recent studies presented some objective metrics to evaluate algorithms. Such metrics consider different ways in which segmentation errors occur (perceptual factors) and its parameters are adjusted according to the application for which the segmented frames are intended. We demonstrate empirically that the performance of existing metrics changes according to the segmentation algorithm; we applied such metrics to evaluate bilayer segmentation algorithms for compose scenes in Augmented Reality environments. We also contribute with a new objective metric to adjust the parameters of two bilayer segmentation algorithms found in the literature.

Video segmentation to extract a person in foreground has been a common task in many Augmented Reality (AR) applications. In natural environments which the background color and the light environment are not constant the segmentation method must be able to extract the element of the interest in these conditions. However, methods for segmentation in natural environment are more error prone than the traditional ones which are based on a constant color elimination. Thus, in order to use them in a large number of applications it is necessary to know how the segmentation errors are perceived by the users in order to focus on development of algorithms which avoid the more perceptible ones. In this work a video quality subjective assessment method was applied to obtain the AR user's opinions about videos with different misclassified pixels rates. The results showed that segmentation errors are perceived by AR applications users. However, the video quality was not related with the number of the misclassified pixels. In addiction, it was noted that when the errors concentrated in the element of the interest increase the score of the associated video decreases.

Augmented Reality (AR) systems which use optical tracking with fiducial marker for registration have had an important role in popularizing this technology, since only a personal computer with a conventional webcam is required. However, in most these applications, the virtual elements are shown only in the foreground a real element does not occlude a virtual one. The method presented enables AR environments based on fiducial markers to support mutual occlusion between a real element and many virtual ones, according to the elements position (depth) in the environment.

This paper presents an algorithm that augments a previous model known in the literature for the automatic segmentation of monocular videos into foreground and background layers. The original model fuses visual cues such as color, contrast, motion and spatial priors within a Conditional Random Field. Our augmented model makes use of bidirectional motion priors by exploiting future evidence. Although our augmented model processes more data, it does so with the same time performance of the original model. We evaluate the augmented model within ground truth data and the results show that the augmented model produces better segmentation.

Most work in navigation approaches for mobile robots does not take into account existing solutions to similar problems when learning a policy to solve a new problem, and consequently solves the current navigation problem from scratch. In this article we investigate a knowledge transfer technique that enables the use of a previously know policy from one or more related source tasks in a new task. Here we represent the knowledge learned as a stochastic abstract policy, which can be induced from a training set given by a set of navigation examples of state-action sequences executed successfully by a robot to achieve a specific goal in a given environment. We propose both a probabilistic and a nondeterministic abstract policy, in order to preserve the occurrence of all actions identified in the inductive process. Experiments carried out attest to the effectiveness and efficiency of our proposal.

The task of navigating to a target position in space is a fairly common task for a mobile robot. It is desirable that this task is performed even in previously unknown environments. One reactive architecture explored before addresses this challenge by denning a hand-coded coordination of primitive behaviours, encoded by the Potential Fields method. Our first approach to improve the performance of this architecture adds a learning step to autonomously find the best way to coordinate primitive behaviours with respect to an arbitrary performance criterion. Because of the limitations presented by the Potential Fields method, especially in relation to non-convex obstacles, we are investigating the use of Relational Reinforcement Learning as a method to not only learn to act in the current environment, but also to generalise prior knowledge to the current environment in order to achieve the goal more quickly in a non-convex structured environment. We show the results of our previous efforts in reaching goal positions along with our current research on generalised approaches.

Usage of real objects as links between real and virtual information is one key aspect in augmented reality. A central issue to achieve this link is the estimation of the visuospatial information of the observed object, or in other words, estimating the object pose. Different objects can have different behaviors when used for interaction. This not only encompasses changes in position, but also folding or deformations. Traditional researches in the area solve those pose estimation problems using different approaches, depending on the type of the object. Additionally, some researches are based only on positional information of observed feature points, simplifying the object information. In this work, we explore the pose estimation of different objects by gathering more information from the observed feature points, and obtaining the local poses of such points, which are not explored in other researches. We apply this local pose estimation idea in two different capturing scenarios, reaching two novel approaches of pose estimation: one based on RGB-D cameras, and another based on RGB and machine learning methods. In the RGB-D based approach, we use the feature point orientation and near surface to obtain its normal; then, find the local 6 degrees-of-freedom (DoF) pose. This approach gives us not only the rigid object pose, but also the approximated pose of deformed objects. On the other hand, our RGB based approach explores machine learning with local appearance changes. Unlike other RGB based works, we replace the complex non-linear systems solvers with a fast and robust method, reaching local rotation of the observed feature points, as well as, full 6 DoF rigid object pose with dramatically lower real-time calculation demands. Both approaches show us that gathering local poses can bring information for the pose estimation of different types of objects. O uso de objetos reais como meio de conexão entre informações reais e virtuais é um aspecto chave dentro da realidade aumentada. Uma questão central para tal conexão é a estimativa de informações visuo-espaciais do objeto, ou em outras palavras, a detecção da pose do objeto. Diferentes objetos podem ter diferentes comportamentos quando utilizados em interações. Não somente incluindo a mudança de posição, mas também sendo dobradas ou deformadas. Pesquisas tradicionais solucionam tais problemas de detecção usando diferentes abordagens, dependendo do tipo de objeto. Adicionalmente, algumas pesquisas se baseiam somente na informação posicional dos pontos de interesse, simplificando a informação do objeto. Neste trabalho, a detecção de pose de diferente objetos é explorada coletando-se mais informações dos pontos de interesse observados e, por sua vez, obtendo as poses locais de tais pontos, poses que não são exploradas em outras pesquisas. Este conceito da detecção de pose locais é aplicada em dois ambientes de capturas, estendendo-se em duas abordagens inovadoras: uma baseada em câmeras RGB-D, e outra baseada em câmeras RGB e métodos de aprendizado de maquinas. Na abordagem baseada em RGB-D, a orientação e superfície ao redor do ponto de interesse são utilizadas para obter a normal do ponto. Através de tais informações a pose local é obtida. Esta abordagem não só permite a obtenção de poses de objetos rígidos, mas também a pose aproximada de objetos deformáveis. Por outro lado, a abordagem baseada em RGB explora o aprendizado de máquina aplicado em alterações das aparências locais. Diferentemente de outros trabalhos baseados em câmeras RGB, esta abordagem substitui solucionadores não lineares complexos com um método rápido e robusto, permitindo a obtenção de rotações locais dos pontos de interesse, assim como, a pose completa (com 6 graus-de-liberdade) de objetos rígidos, com uma demanda computacional muito menor para cálculos em tempo-real. Ambas as abordagens mostram que a coleta de poses locais podem gerar informações para a detecção de poses de diferentes tipos de objetos.