André F. T. Martins

Unbabel's Participation in the WMT16 Word-Level Translation Quality Estimation Shared Task.

This paper presents the contribution of the Unbabel team to the WMT 2016 Shared Task on Word-Level Translation Quality Estimation. We describe our two submitted systems: (i) UNBABELLINEAR, a feature-rich sequential linear model with syntactic features, and (ii) UNBABEL-ENSEMBLE, a stacked combination of the linear system with three different deep neural networks, mixing feedforward, convolutional, and recurrent layers. Our systems achieved F OK 1 × F BAD 1 scores of 46.29% and 49.52%, respectively, which were the two highest scores in the challenge.

Combining free energy score spaces with information theoretic kernels: Application to scene classification

Most approaches to learn classifiers for structured objects (e.g., images) use generative models in a classical Bayesian framework. However, state-of-the-art classifiers for vectorial data (e.g., support vector machines) are learned discriminatively. A generative embedding is a mapping from the object space into a fixed dimensional score space, induced by a generative model, usually learned from data. The fixed dimensionality of these generative score spaces makes them adequate for discriminative learning of classifiers, thus bringing together the best of the discriminative and generative paradigms. In particular, it was recently shown that this hybrid approach outperforms a classifier obtained directly for the generative model upon which the score space was built. Using a generative embedding involves two steps: (i) defining and learning the generative model and using it to build the embedding; (ii) discriminatively learning a (maybe kernel) classifier on the adopted score space. The literature on generative embeddings is essentially focused on step (i), usually using some standard off-the-shelf tool for step (ii). In this paper, we adopt a different approach, by focusing also on the discriminative learning step. In particular, we combine two very recent and top performing tools in each of the steps: (i) the free energy score space; (ii) non-extensive information theoretic kernels. In this paper, we apply this methodology in scene recognition. Experimental results on two benchmark datasets shows that our approach yields state-of-the-art performance.

Information theoretical Kernels for generative embeddings based on hidden Markov models

Many approaches to learning classifiers for structured objects (e.g., shapes) use generative models in a Bayesian framework. However, state-of-the-art classifiers for vectorial data (e.g., support vector machines) are learned discriminatively. A generative embedding is a mapping from the object space into a fixed dimensional feature space, induced by a generative model which is usually learned from data. The fixed dimensionality of these feature spaces permits the use of state of the art discriminative machines based on vectorial representations, thus bringing together the best of the discriminative and generative paradigms. n nUsing a generative embedding involves two steps: (i) defining and learning the generative model used to build the embedding; (ii) discriminatively learning a (maybe kernel) classifier on the adopted feature space. The literature on generative embeddings is essentially focused on step (i), usually adopting some standard off-the-shelf tool (e.g., an SVM with a linear or RBF kernel) for step (ii). In this paper, we follow a different route, by combining several HiddenMarkov Models-based generative embeddings (including the classical Fisher score) with the recently proposed non-extensive information theoretic kernels. We test this methodology on a 2D shape recognition task, showing that the proposed method is competitive with the state-of-art.

Combining information theoretic kernels with generative embeddings for classification

Classical approaches to learn classifiers for structured objects (e.g., images, sequences) use generative models in a standard Bayesian framework. To exploit the state-of-the-art performance of discriminative learning, while also taking advantage of generative models of the data, generative embeddings have been recently proposed as a way of building hybrid discriminative/generative approaches. A generative embedding is a mapping, induced by a generative model (usually learned from data), from the object space into a fixed dimensional space, adequate for discriminative classifier learning. Generative embeddings have been shown to often outperform the classifiers obtained directly from the generative models upon which they are built. Using a generative embedding for classification involves two main steps: (i) defining and learning a generative model and using it to build the embedding; (ii) discriminatively learning a (maybe kernel) classifier with the embedded data. The literature on generative embeddings is essentially focused on step (i), usually taking some standard off-the-shelf tool for step (ii). Here, we adopt a different approach, by focusing also on the discriminative learning step. In particular, we exploit the probabilistic nature of generative embeddings, by using kernels defined on probability measures; in particular we investigate the use of a recent family of non-extensive information theoretic kernels on the top of different generative embeddings. We show, in different medical applications that the approach yields state-of-the-art performance.

Renal cancer cell classification using generative embeddings and information theoretic kernels

In this paper, we propose a hybrid generative/discriminative classification scheme and apply it to the detection of renal cell carcinoma (RCC) on tissue microarray (TMA) images. In particular we use probabilistic latent semantic analysis (pLSA) as a generative model to perform generative embedding onto the free energy score space (FESS). Subsequently, we use information theoretic kernels on these embeddings to build a kernel based classifier on the FESS. We compare our results with support vector machines based on standard linear kernels and RBF kernels; and with the nearest neighbor (NN) classifier based on the Mahalanobis distance using a diagonal covariance matrix. We conclude that the proposed hybrid approach achieves higher accuracy, revealing itself as a promising approach for this class of problems.

An Integrated Evaluation Method for E-Learning: A Case Study.

Technological improvements in broadband and distributed computing are making it possible to distribute live media content cost‐effectively. Because of this, organizations are looking into cost‐effective approaches to implement e‐Learning initiatives. Indeed, computing resources are not enough by themselves to promote better e‐Learning experiences. Hence, our goal is to share preliminary results on testing a holistic evaluation method for e‐Learning environments. To this end, we have built an experience within class dynamics using an open source Learning Virtual Environment integrated with webcast and video archive features. Our proposed evaluation method capyures user feedback by classifying it according to motivation to e‐learn in groups, since we have found this approach simpler than using classic behavioural methods. This helped us to define practical design guidelines to yield faster and more efficient e‐Learning development processes. Our results show that consistent communication both online and offline, translates into efficiency. It also dampens negative perceptions during the transition from traditional to online learning environments. These results will contribute in designing intervention strategies to optimize organizational investments in e‐Learning across user groups and contexts.

Navigating in Manhattan: 3D orientation from video without correspondences

The problem of inferring 3D orientation of a camera from video sequences has been mostly addressed by first computing correspondences of image features. This intermediate step is now seen as the main bottleneck of those approaches. In this paper, we propose a new 3D orientation estimation method for urban (indoor and outdoor) environments, which avoids correspondences between frames. The basic scene property exploited by our method is that many edges are oriented along three orthogonal directions; this is the recently introduced Manhattan world (MW) assumption. In addition to the novel adoption of the MW assumption for video analysis, we introduce the small rotation (SR) assumption, that expresses the fact that the video camera undergoes a smooth 3D motion. Using these two assumptions, we build a probabilistic estimation approach. We demonstrate the performance of our method using real video sequences.

Self-Diagnosing Low Coverage and High Interference in 3G/4G Radio Access Networks based on Automatic RF Measurement Extraction

This paper presents a new approach for automatic detection of low coverage and high interference scenarios n n(overshooting and pilot pollution) in Universal Mobile Telecommunications System (UMTS) /Long Term Evolution n n(LTE) networks. These algorithms, based on periodically extracted Drive Test (DT) measurements (or n nnetwork trace information), identify the problematic cluster locations and compute harshness metrics, at cluster n nand cell level, quantifying the extent of the problem. Future work is in motion by adding self-optimization n ncapabilities to the algorithms, which will automatically suggest physical and parameter optimization actions, n nbased on the already developed harshness metrics. The proposed algorithms were validated for a live network n nurban scenario. 830 3rd Generation (3G) cells were self-diagnosed and performance metrics were computed. n nThe most negative detected behaviors regards high interference control and not coverage verification.

On the Combination of Information-Theoretic Kernels with Generative Embeddings

Classical methods to obtain classifiers for structured objects (e.g., sequences, images) are based on generative models and adopt a classical generative Bayesian framework. To embrace discriminative approaches (namely, support vector machines), the objects have to be mapped/embedded onto a Hilbert space; one way that has been proposed to carry out such an embedding is via generative models (maybe learned from data). This type of hybrid discriminative/generative approach has been recently shown to outperform classifiers obtained directly from the generative model upon which the embedding is built.

A Hybrid Neighbor Optimization Algorithm for SON based on Network Topology, Handover Counters and RF Measurements

With the increasing complexity of current wireless networks, it became evident the need for Self-OrganizingnNetworks (SON), which aims to automate most of the associated radio planning and optimization tasks. WithinnSON, this paper aims to optimize the Neighbor Cell List (NCL) for radio network cells. An algorithm composednby three decision criteria was developed: geographic localization and orientation, according networkntopology, Radio Frequency (RF) measurements collected by drive-tests or traces and Performance Managementn(PM) counters from Handover (HO) statistics. The first decision, proposes a new NCL taking intonaccount the Base Station (BS) location and interference tiers, based on the quadrant method. The last twondecision criteria consider signal strength and interference level measurements and HO statistics in a time period,nrespectively. They also define a priority to each cell and added, kept or removed neighbor relation,nbased on user defined constraints. The algorithms were developed and implemented over new radio networknoptimization professional tool. Several case studies were produced using real data from a mobile operator.