posters-2

Multiagent reinforcement learning has shown its potential for tackling real world problems, like traffic. We consider the toll-based route choice problem, where self-interested drivers need to repeatedly choose routes that minimise their travel times. A major challenges here is to deal with agents' selfishness when competing for a common resource, as they tend to converge to a substantially far-from-optimum equilibrium. In traffic, this translates into higher congestion levels. Road tolls have been advocated as a means to tackle this issue, though typically assuming that (i) drivers have homogeneous preferences, and that (ii) collected tolls are kept for the traffic authority. In this paper, we propose Generalised Toll-based Q-learning (GTQ-learning), a multiagent reinforcement learning algorithm capable of realigning agents' heterogeneous preferences with respect to travel time and monetary expenses. GTQ-learning neutralises agents' preferences, thus ensuring that congestion levels are minimised regardless of agents' selfishness levels. Furthermore, GTQ-learning achieves approximated budget balance by redistributing a fraction of the collected tolls. We perform a theoretical analysis of GTQ-learning, showing that it leads agents to a system-efficient equilibrium, and provide empirical results, evidencing that GTQ-learning minimises congestion on realistic road networks.

Optimization of hyper-parameters in reinforcement learning (RL) algorithms is a key task, because they determine how the agent will learn its policy by interacting with its environment, and thus what data is gathered. In this work, an approach that uses Bayesian optimization to perform a two-step optimization is proposed: first, categorical RL structure hyper-parameters are taken as binary variables and optimized with an acquisition function tailored for such variables. Then, at a lower level of abstraction, solution-level hyper-parameters are optimized by resorting to the expected improvement acquisition function, while using the best categorical hyper-parameters found in the optimization at the upper-level of abstraction. This two-tier approach is validated in a simulated task.

Machine learning theory has mostly focused on generalization to samples from the same distribution as the training data. Whereas a better understanding of generalization beyond the training distribution where the observed distribution changes is also fundamentally important to achieve a more powerful form of generalization. In this paper, we attempt to study through the lens of information measures how a particular architecture behaves when the true probability law of the samples is potentially different at training and testing times. Our main result is that the testing gap between the empirical cross-entropy and its statistical expectation (measured with respect to the testing probability law) can be bounded with high probability by the mutual information between the input testing samples and the corresponding representations, generated by the encoder obtained at training time. These results of theoretical nature are supported by numerical simulations showing that the mentioned mutual information is representative of the testing gap, capturing qualitatively the dynamic in terms of the hyperparameters of the network.

"We introduce Post-DAE, a post-processing method based on denoising autoencoders to improve the anatomical plausibility of arbitrary biomedical image segmentation algorithms. Some of the most popular segmentation methods still rely on post-processing strategies like conditional random fields to incorporate connectivity constraints into the resulting masks. Even if it is a valid assumption in general, these methods do not offer a straightforward way to incorporate more complex priors like convexity or arbitrary shape restrictions. Post-DAE leverages the latest developments in manifold learning via denoising autoencoders. We learn a low-dimensional space of anatomically plausible segmentations, and use it to impose shape constraints by post-processing anatomical segmentation masks obtained with arbitrary methods. Our approach is independent of image modality and intensity information since it employs only segmentation masks for training. We performed experiments in segmentation of chest X-ray images. Our experimental results show that Post-DAE can improve the quality of noisy and incorrect segmentation masks obtained with a variety of standard methods, by bringing them back to a feasible space, with almost no extra computational cost."

El trabajo busca plantear el modelo de un sistema asistente de velocidad vehicular que sea capaz de identificar coordenadas, en el vehículo y compararlas con las coordenadas almacenadas, por medio de una webApp implementando API’S de google maps, en una base de datos de marcas de velocidad de zonas urbanas, y alertar al conductor si excede en alguna de ellas a medida que vaya avanzando en su recorrido, lo que puede contribuir de gran manera al desarrollo de esta línea de investigación y a mejorar de forma notable las posibilidades de implementar en un futuro no muy lejano en un automóvil que pueda ser completamente asistido por un computador, teniendo en cuenta estos principios aquí mencionados.

"The high efficiency of energy transfer is one of the main motivations in the study of light-harvesting systems. The accurate description of these complexes can be formulated in the framework of open quantum systems which comprises the interaction among their fundamental units called chromophores and the interaction with the environment. Maximizing energy transfer involves optimally controlled system dynamics and at the same time, getting optimal configurations that achieve this objective. Therefore, this research proposes the implementation of reinforcement learning (RL) as a mechanism for quantum optimal control of excitation energy transfer (EET) in light-harvesting systems and, in turn, obtaining configurations that maximize efficiency through a classical agent that even can tolerate environments with high noise levels. "

"Recently, several machine learning based models have been proposed for Natural Language Processing (NLP), achieving outstanding results, by using powerful architectures like “Transformer” (Vaswani et al., 2017) and pretraining on large text corpus, as is the case of BERT (Devlin et al., 2018). However, it has been shown that language models are fragile (they are easily broken) and biased (instead of an actual comprehension of the text, they tend to take advantage of data biases). To the best of our knowledge, this is the first time a Transformer-based model is systematically put to test."

Molecular data from cancer patients is characterized by tens of thousands of gene features and also by different modalities or ‘omics’ like Genomics, Transcriptomics and Proteomics. These systems are also labeled by clinical information like patient survival, tumor stage and tumor subtype. The initial high dimensional input space is noisy and makes complicated to find useful patterns like similarities between tumor types and sub-types. For clinical reasons this work aims to learn meaningful and lower dimensional representations of tumors which keeps biological signals and contribute to classify tumor subtype or stage by using Variational Autoencoders (VAE) and Kernelized Autoencoders (KAE) . Then a feature selection strategy based on Multiple Kernel Learning is executed with the objective to approximate as much as possible the resulting representation based on the selected features to the one learned by the autoencoders. Selected features are also evaluated to classify tumor samples based on clinical labels and also to discover tumor subtypes. Preliminary results show that the learned representations drive the selection of meaningful genes associated to the clinical outcome of the patient and thus provide evidence for potential biomarkers.

Despite of the recent progress in agents that learn through interaction, there are several challenges in terms of sample efficiency and generalization across unseen behaviors during training. To mitigate these problems, we propose and apply a first-order Meta-Learning algorithm called Bottom-Up Meta-Policy Search (BUMPS), which works with two-phase optimization procedure: firstly, in a meta-training phase, it distills few expert policies to create a meta-policy capable of generalizing knowledge to unseen tasks during training; secondly, it applies a fast adaptation strategy named Policy Filtering, which evaluates few policies sampled from the meta-policy distribution and selects which best solves the task. We conducted all experiments in the RoboCup 3D Soccer Simulation domain, in the context of kick motion learning. We show that, given our experimental setup, BUMPS works in scenarios where simple multi-task Reinforcement Learning does not. Finally, we performed experiments in a way to evaluate each component of the algorithm.

The Classification of regions, especially urban areas, on synthetic aperture polarimetric radar (PolSAR) data is a challenging task. We know that texture analysis has a great informational power of the spatial properties of the main elements of the image, being one of the most important techniques in image processing and pattern recognition. The first task of this analysis is the extraction of discriminant features capable of efficiently incorporating information about the characteristics of the original image. Based on this principle, in this paper, we propose a new classification technique. Through the analysis of the textures of these images, ordinal pattern transition graphs, and information theory descriptors, we achieved a high discriminatory power in the characterization and classification of the regions under study.

The fluctuations in the temperature have a strong influence in the electric consumption. As a consequence, identifying and finding groups of possible climate scenarios is useful for the analysis of the electric supply system. The scenarios data that we are considering are time series of hourly measured temperatures over a grid of geographical points in France and neighboring areas, used by the French company RTE . Clustering techniques are useful for finding homogeneous groups of times series but the challenge is to find a suitable data transformation and distance metric. In this work, we used several transformations (fourier, wavelets, autoencoders) and distance metrics (DTW and euclidean among others) and found consistent groups of climate scenarios using clustering techniques. We give several performance indicators and we findd that k-shape performs the best according to some of them.

Considering the occurrence of an event with high social impact, it is important to establish a space-time relation of available information and so, answer some questions about the event as“who”, “how”, “where” and “why”. This work is part of the thematic FAPESP project “DéjàVu: Feature-Space-Time Coherence from Heterogeneous Data for Media Integrity Analytics and Interpretation of Events” and it proposes, from social network collected data, to determine the relevance of them for the analyzed event, allowing the correct construction of relationships among these data during an analysis phase later on. The main challenges of this work are the characteristics of the data which will be used: heterogeneity, as they come from different sources; multi-modality, such as texts, images and videos; unlabeled data, as they do not present label of straightforward relevance for the event; and unstructured data, as they do not possess characteristics which could be used directly during the learning.

"We introduce Concept Attention Networks (CAN) for Action Recognition. CANs seek to provide more interpretability by providing attention for both visual features as well as concepts associated to the action we want to recognize. CANs are modelled on the MAC architecture - which has produced great results on VQA through the use of sequential reasoning- with two main differences: 1) The knowledge base is modified to take video features. 2) We introduce attention over concepts via an auxiliary task that tries to guess the concepts associated to the predicted class on each reasoning step. We expect that taking visual features and word features to the same space might provide both similar accuracy as well as greater interpretability; since CAN - as the MAC architecture on which it is based- divides its reasoning in steps, we are able to see on which parts of the video and on which concepts the model is focusing to generate its predictions. We present results on the Something to Something v2 dataset against a C3D baseline. "

"The work consists in applying Deep Reinforcement Learning algorithms in order to the improve a robot's walking engine. Therefore, the final goal is to implement a Push Recovery Controller, which is a bio-inspired controller that stabilizes the agent under external perturbations in order to achieve a more stable and also faster walking movement. Proximal Policy Optimization algorithm has already been used in different domains and had success to solve many Continuous Control problems, being considered one of the state-of-art techniques of Deep Reinforcement Learning, therefore this is the main technique used in this work. Given the nature of Policy Gradient methods, we applied distributed training in order to Speed Up the learning process. We have used Intel AI DevCloud Cluster in order to have many agents running in parallel."

The application of deep learning to symbolic domains remains an active research endeavour. Graph neural networks (GNN), consisting of trained neural modules which can be arranged in different topologies at run time, are sound alternatives to tackle relational problems which lend themselves to graph representations. In this paper, we show that GNNs are capable of multitask learning, which can be naturally enforced by training the model to refine a single set of multidimensional embeddings and decode them into multiple outputs by connecting MLPs at the end of the pipeline. We demonstrate the multitask learning capability of the model in the relevant relational problem of estimating network centrality measures, focusing primarily on producing rankings based on these measures. We then show that a GNN can be trained to develop a lingua franca of vertex embeddings from which all relevant information about any of the trained centrality measures can be decoded. The proposed model achieves 89% accuracy on a test dataset of random instances with up to 128 vertices and is shown to generalise to larger problem sizes. The model is also shown to obtain reasonable accuracy on a dataset of real world instances with up to 4k vertices, vastly surpassing the sizes of the largest instances with which the model was trained ($n=128$). Finally, we believe that our contributions attest to the potential of GNNs in symbolic domains in general and in relational learning in particular.

Recent research has shown the potential of learning lane following policies from annotated video sequences through the use of advanced machine learning techniques. They however require high computational power, prohibiting their use in low-budget projects such as educational robotic kits and embedded devices. Sum-product networks (SPNs) are a class of deep probabilistic models with clear probabilistic semantics and competitive performance. Importantly, SPNs learned from data are usually several times smaller than deep neural networks trained for the same task. In this work, we develop an end-to-end imitation learning solution to lane following using SPNs to classify images into a finite set of actions. Images are obtained from a monocular camera, which is part of the low-cost custom made mobile robot. Our results show that our solution generalizes training conditions with relatively few data. We investigate the trade-off between computational and predictive performance, and conclude that sacrificing accuracy for the benefit of faster inference results in improved performance in the real world, especially in resource constrained environments.

" The social Web, mainly represented by social networking sites, enriches the life and activities of its users by providing new forms of communication and interaction. Even though most of the time, the use of Internet is safe and enjoyable, there are risks that involve communication through social media. The unmoderated nature of social media sites often results in the appearance and distribution of unwanted content or misinformation. Thus, although social sites provide a great opportunity to stay informed about events and news, it also produces skepticism among users, as not every piece of shared information can be trusted. Moreover, the potential for automation and the low cost of producing fraudulent sites, allows the rapid creation and dissemination of unwanted content. Thus, current information dissemination processes pose the challenge of determining whether it is possible to trust on recommendations. The goal of this work is to define a profile to describe and estimate the trustworthiness or reputation of users, to avoid making recommendations that could favour the propagation of unreliable content and polluting users. The final aim is to reduce the negative effects of the existence and propagation of such content, and thus improving the quality of the recommendations."

"Sum-Product Networks (SPN) are deep probabilistic models that have exhibited state-of-the-art performance in several machine learning tasks. As with many other probabilistic models, performing Maximum-A-Posteriori (MAP) inference is NP-hard in SPNs. A notable exception is selective SPNs, that allows MAP inference in linear time. Due to the high number of parameters, SPNs learned from data can produce unreliable and overconfident inference. This effect can be partially detected by performing a Sensitivity Analysis of the model predictions to changes in the parameters. In this work, we develop efficient algorithms for global quantitative analysis of MAP inference in selective SPNs. In particular, we devise a polynomial-time procedure to decide whether a given MAP configuration is robust with respect to changes in the model parameters. Experiments with real-world datasets show that this approach can discriminate easy- and hard-to-classify instances, often more accurately than criteria based on the probabilities induced by the model."

"In recent years, the amount of available data has drastically increased. However, labelling such data is hugely expensive. In this scenario, semi-supervised learning emerge as a vitally important tool, which combines labelled data (supervised machine learning) and unlabelled data (unsupervised learning) in order to make better predictions. In particular, graph based algorithms takes into account the relationships between the instances of the data and the underlying graph structures to make those predictions. In addition, in the context of data analysis, there are scenarios that can be naturally think as graphs. This occurs in situations where in addition to individual properties, connectivity between the elements of the data set is also important. Therefore, it is logical that machine learning models include information from both a node and its neighbours when making a prediction. This works propose adding graph feature regularization terms (GFR) to the the objective function to maximize. This new regularization terms depends on the structure of the network, the weight of the edges and the features of the node. We conclude that adding this terms to gradient boosted trees can outperform complex network architectures such as the Graph Convolutional Networks."

The convergence of High-Performance Computing (HPC) and Artificial Intelligence (AI) has become a promissing approach to major performance improvements. This combination has much to offer from each other and it’s giving to the users unprecedent capabilities of research. In this interaction HPC could be used by AI (HPC for AI) to execute and enhance the performance of its algorithms. It involves using and evaluating different HPC architectures to train AI algorithms, understand and optimize their performance on different architectures. IA for HPC can be further subdivided in IA after HPC and autotune. In the first, ML algorithms are used to understand and analyze the results of simulations on HPC. It involves using ML to understand scientific applications, how they are relate to different HPC architectures, and the impact of this relationship on performance and power consumption. It is more related to knowledge discovery and its result can be used in autotune. In autotune, IA is used to configure HPC, to choose the best set of computation and parameters to achieve some goal, for example energy saving. Also, ML is used to the prediction of performance and energy consumption, job scheduling and frequency and voltage scaling.

A class of algorithms known as feature least-mean-square (F-LMS) has been proposed recently to exploit hidden sparsity in adaptive filter parameters. In contrast to common sparsity-aware adaptive filtering algorithms, the F-LMS algorithm detects and exploits sparsity in linear combinations of filter coefficients. Indeed, by applying a feature matrix to the adaptive filter coefficients vector, the F-LMS algorithm can revealand exploit their hidden sparsity. However, in many cases the unknown plant to be identified contains not only hidden but also plain sparsity and the F-LMS algorithm is unable to exploit it. Therefore, we can incorporate sparsity-promoting techniques into the F-LMS algorithm in order to allow the exploitation of plain sparsity. In this paper, by utilizing the l0-norm, we propose the l0-norm F-LMS (l0-F-LMS) algorithm for sparse lowpass and sparse highpass systems. Numerical results show that the proposed algorithm outperforms the F-LMS algorithm when dealing with hidden sparsity, particularly in highly sparse systems where the convergence rate is sped up significantly.

Graph Neural Networks are a promising technique for bridging differential programming with combinatorial domains. In this paper we show that GNNs can learn to solve, with very little supervision, the decision variant of the Traveling Salesperson Problem.

Recent advances in the field of Deep Reinforcement Learning have achieved impressive results in various tasks. One key component for these achievements are the simulated environments used to train and test DeepRL based agents, and for locomotion tasks there are various benchmarks that can be used, which are built on top of popular physics engines. However, these locomotion benchmarks do not offer the functionality required to train and evaluate agents in more diverse and complex tasks, exposing only relatively simple tasks, e.g. traversing flat terrain. This work presents an engine-agnostic toolkit for locomotion tasks that provides such functionality, allowing users to create a wide range of diverse and complex environments. We provide support for various physics engines via a physics abstraction layer, allowing users to easily switch between engines as required.

Optimizations at placement stage need to be guided by timing estimation prior to routing. To handle timing uncertainty due to the lack of routing information, people tend to make very pessimistic predictions such that performance specification can be ensured in the worst case. Such pessimism causes over-design that wastes chip resources or design effort. In this work, a machine learning-based pre-routing timing prediction approach is introduced. Experimental results show that it can reach accuracy near post-routing sign-off analysis. Compared to a commercial pre-routing timing estimation tool, it reduces false positive rate by about 2/3 in reporting timing violations.

Memory is an important aspect of intelligence and plays a role in many deep reinforcement learning models. However, little progress has been made in understanding when specific memory systems help more than others and how well they generalize. The field also has yet to see a prevalent consistent and rigorous approach for evaluating agent performance on holdout data. In this paper, we aim to develop a comprehensive methodology to test different kinds of memory in anagent and assess how well the agent can apply what it learns in training to a holdout set that differs from the training set along dimensions that we suggest are relevant for evaluating memory-specific generalization. To that end, we first construct a diverse set of memory tasks that allow us to evaluate test-time generalization across multiple dimensions. Second, we develop and perform multiple ablations on an agent architecture that combines multiple memory systems, observe its baseline models, and investigate its performance against the task suite.

Industrial applications, such as those in process industry like or power generation, could benefit from reinforcement learning (RL) agents to reduce energy consumption and lower emissions. However, the systems involved in these applications usually have high usage costs, while RL algorithms generally require too many trials to learn a task. A promising approach to the inefficiency problem is the model-based RL method, which allows agents to learn a predictive model of the environment to extract more information from available data. Given that industrial applications generally feature complex stochastic behavior, we propose investigating novel integration schemes between the model-based approach and deep generative models, a class of neural networks specially designed to handle sophisticated probability distributions. We will test these interventions in existing and novel benchmark tasks aimed at assessing a learning system's capacity of handling state changes governed by complex conditional probability distributions. We expect that our approach will lead to better model predictions and faster learning.

Tikhonov functionals are commonly used as regularization strategies for severely ill-posed inverse problems. Besides the type of penalization induced into the solution, the proper selection of the regularization parameters is of utmost importance for accurate estimation. In this work, we analyze several data-driven regularization parameters estimation methods in a mixed-term discriminative framework. Numerical results for P300 detection in Brain-Computer Interfaces classification are presented, showing the impact of regularization parameter estimation into classification performance.

"The number, variety, and complexity of Data Science applications are rapidly increasing along with automated solutions. This kind of solution, called automated machine learning, makes data science accessible to non-specialists. On the other hand, from the specialist standpoint, automated machine learning can spare him/her manual and repetitive work, speeding up research. In the last years, there has been a strong interest in the development of tools able to automate data science. While the existing frameworks mainly focus on inducing accurate models through hyperparameter tuning, they disregard or forgo, for instance, the data preprocessing step, reproducibility, and explainability. Nevertheless, this kind of task expends the majority of human resources. In this paper, we present an overview of ideas behind Pajé, an open tool for automated data science. Pajé includes all the core processes of the data science pipeline, from data acquisition to model interpretation, and at the same time, addresses important aspects of machine learning, such as reproducibility and explainability."

"Proteins are the most versatile biological molecules, with diverse functions. Recently, the AI community have developed interest in specific topics related to proteins as: protein folding, structural analysis, protein-ligand affinity estimation, among others. Cruzain is a cysteine protease involved in chagas disease with several Cz-inhibitor complexes deposited in the Protein Data Bank (PDB). Unfortunately, the number of structures solved up-to-date is scarce for the requirements of a machine learning optimization algorithm. Another issue is the high dimensionality of the data involved in structure-based approaches for drug design. In this work, charge density-based data was employed as input for a classification algorithm with the protein-ligand interactions as columns and ligands as rows. A support vector machine with recursive feature elimination was employed to uncover the most relevant features involved in the protein-inhibitor complexes. This approach is the first step for further analysis of topological data of Cz-ligand complexes under study. We hope that results will shed light to understand the inhibition mechanism of Cruzain."

We introduce a cutting plane method for efficiently finding the probability distribution of maximum entropy contained in a Wasserstein ball. Such distributions are the most general (i.e. minimizers of the amount of prior information) in the ball and are therefore of central importance for statistical inference. We generalize these results to the problem of minimizing cross-entropy from a given prior distribution and use them to propose 1-parameter families of learning algorithms that are naturally resilient to biases.

We automatically design kernels from data by projecting points into either random hyperplanes or onto the boundaries forming the $\alpha$-shape. We interpret such transformation as an explicit strategy a kernel uses to extract features from data, thus SVM applied on this transformed space should be capable of correctly separating class instances. We firstly applied this method on two different synthetic datasets to assess its performance and parameter sensitivity. Those experimental results confirmed a considerable improvement over the original input space, robustness in the presence of noise and parameter changes. Secondly, we applied our approach on well-known image datasets in order to evaluate its ability to deal with real-world data and high dimensional spaces. Afterwards, we discuss how this novel approach could be plugged to Convolutional Neural Networks, helping to understand the effects and the impact of adding units to layers. Our proposal has a low computational cost and it is parallelizable to work directly on the transformed space and, when memory constraints hold, its resultant kernel matrix might be used instead. Such approach considerably improved the classification performance in almost all scenarios, supporting the claim that it could be used as a general-purpose kernel transformation.

This work explores the general problem of explaining the behavior of recurrent neural networks (RNN). The goal is to construct a representation which enhances human understanding of an RNN as a sequence classifier, with the purpose of providing insight on the rationale behind the classification of a sequence as positive or negative, but also to enable performing further analyses, such as automata-theoretic formal verification. In particular, an active learning algorithm for constructing a deterministic finite automaton which is approximately correct with respect to an artificial neural network is proposed.

We present new algorithms for computing and approximating bisimulation metrics in Markov Decision Processes (MDPs). Bisimulation metrics are an elegant formalism that capture behavioral equivalence between states and provide strong theoretical guarantees on differences in optimal behaviour. Unfortunately, their computation is expensive and requires a tabular representation of the states, which has thus far rendered them impractical for large problems. In this paper we present a new version of the metric that is tied to a behavior policy in an MDP, along with an analysis of its theoretical properties. We then present two new algorithms for approximating bisimulation metrics in large, deterministic MDPs. The first does so via sampling and is guaranteed to converge to the true metric. The second is a differentiable loss which allows us to learn an approximation even for continuous state MDPs, which prior to this work had not been possible.

Mental disorders such as depression and anxiety have been increasing at alarming rates in the worldwide population. Notably, major depressive disorder has become a common problem among higher education students. While the reasons for this alarming situation remain unclear (although widely investigated), the student already facing this problem must receive treatment. To that, it is first necessary to screen the symptoms. The traditional way for that is relying on clinical consultations or answering questionnaires. However, nowadays, the data shared at social media is a ubiquitous source that can be used to detect the depression symptoms even when the student is not able to afford or search for professional care. In this work, we focus on detecting the severity of the depression symptoms in higher education students, by comparing deep learning with feature engineering models induced from Instagram data. The experimental results show that students presenting a BDI score higher than 20 can be detected with 0.92 of recall and 0.69 of precision in the best case, reached by a fusion model. Our findings show a potential of help on further investigation of depression, by bringing students at-risk to light, to guide them to access adequate treatment.

"We address the problem of solving ill-posed inverse problems in imaging where the prior is a neural generative model. Specifically we consider the decoupled case where the prior is trained once and can be reused for many different degradation models without retraining. Whereas previous MAP-based approaches to this problem lead to highly non-convex optimization algorithms, our approach computes the joint (space-latent) MAP that naturally leads to alternate optimization algorithms and to the use of a stochastic encoder to accelerate computations. We show theoretical and experimental evidence that the proposed objective function may be quite close to bi-convex, which would pave the way to show strong convergence results of our optimization scheme. Experimental results also show the higher quality of the solutions obtained by our approach with respect to non-convex MAP approaches."

"The application of machine learning to Network Measurements Analysis problems has largely increased in the last decade; however, it remains difficult to say today which is the most fitted category of models to address these tasks in operational networks. We work on Stream-GML2, a generic stream-based (online) Machine Learning model for the analysis of network measurements. The model is a stacking ensemble learning algorithm, in which several weak or base learning algorithms are combined to obtain higher predictive performance. In particular, Stream-GML2 is an instance of a recent model known as Super Learner, which performs asymptotically as good as the best input base learner. It provides a very powerful approach to tackle multiple problems with the same technique while minimizing over-fitting likelihood during training, using a variant of cross-validation. Additionally, stream-GML2 copes with concept drift and performance degradation by relying on Reinforcement Learning (RL) principles, no-regret learning and online-convex optimization. The model resorts to adaptive memory sizing to retrain the system when required, adjusting its operation point dynamically according to distribution changes in incoming samples or performance degradation over time."

We present a novel application to infer atmospheric radar reflectivity images using infrared satellite images. Given the high cost of radar instruments, data oriented image reconstruction appears as an attractive option. We compared output from fully connected networks, convolutional-deconvolutional networks and generative adversarial networks trained with synthetically generated radar/satelite image pairs from numerical weather model simulations. Results are comparable with state of the art statistical methods. The application shows promising results for short term weather prediction.

Legacy IP networks are currently a huge problem for Internet Service Providers, as the demand grows exponentially, the profit doesn't follow. With the emergence of the Software-Defined Networks (SDN), providers are hoping to improve their service while lowing the operational expenses. In this work, we focus on self-healing SDNs, that requires fault-tolerant mechanisms and intelligent network management for enabling the system to perceive its incorrect states and acting to fix it. As fault tolerance is a huge issue, we narrow our proposal for only dynamic failures, as these are usually the best target for machine learning approaches as deterministic solutions are sub-optimal or too complex. Thus, we develop a solution using Deep Reinforcement Learning (DRL) for routing and load balancing, considering highly dynamic traffic, and we show the viability of a model-free solution and its efficiency.

As in other verticals, the application of data science to education opens up new possibilities. An example is the growing research community in learning analytics. Different goals, such as looking for tools for a more personalized education and the detection of particular difficulties at early ages, are relevant challenges that are being addressed in the area. In this context, we present the case of Plan Ceibal, an institution that assists the education system in Uruguay, providing technological solutions for the support of education.

Reinforcement learning methods have successfully been applied to build autonomous agents that solve challenging sequential decision-making problems. However, agents need a long time to learn a task, especially when multiple autonomous agents are in the environment. This research aims to propose a Transfer Learning framework to accelerate learning by combining two knowledge sources: (i) previously learned tasks; and (ii) advice from a more experienced agent. The definition of such a framework requires answering several challenging research questions, including: How to abstract and represent knowledge, in order to allow generalization and posterior reuse?, How and when to transfer and receive knowledge in an efficient manner?, and How to consistently combine knowledge from several sources?

"Alternatives to recurrent neural networks, in particular, architectures based on attention, have been gaining momentum for processing input sequences. In spite of their relevance, the computational properties of these alternatives have not yet been fully explored. We study the computational power of one of the most paradigmatic architectures exemplifying the attention mechanism, the Transformer (Vaswani et al., 2017). We show that the Transformer is Turing complete exclusively based on their capacity to compute and access internal dense representations of the data. Our study also reveals some minimal sets of elements needed to obtain these completeness results."

Many branches of science and engineering require differential equations to model the dynamics of the systems under study. Traditionally, the identification of the appropriate terms in the equation has been done by experts. Brunton, Proctor, and Kutz 2016 developed a method to automate this task using the data itself. In this work, we extend the applicability of this method to situations where not all variables are observed by adding higher-order derivatives to the model space search. We first test the results using only one variable of the Lorenz system and then apply the same methodology to temperature times series. We found that the proposed approach is enough to recover equations with R²>0.95 in both cases. We also propose an algebraic method to get future values of the system and compare it with traditional integrative methods finding that our approach is more stable giving high accuracy prediction results in the case of the Lorenz system.

We present an integrated interactive framework for the visual analysis of time-varying multivariate datasets. As part of our research, we performed in-depth studies concerning the applicability of visualization techniques to obtain valuable insights. We consolidated the considered analysis and visualization methods in one framework, called TV-MV Analytics. It effectively combines visualization and data mining algorithms providing the following capabilities: i) visual exploration of multivariate data at different temporal scales; and ii) a hierarchical small multiples visualization combined with interactive clustering and multidimensional projection to detect temporal relationships in the data. We demonstrate the value of our framework for specific scenarios, by studying three use cases that were validated and discussed with domain experts.

dSense is a specialized R&D Studio that provides consultancy and development services in Computer Vision, Machine Learning and Image Processing for projects with an important component of innovation. Our team of 4 PhDs, 5 MScs and experienced engineers authored more than 175 papers and 4 US patents. By taking advantage of our research background, we have been able to develop valuable custom AI-enabled solutions across industries with a positive social and productivity impact. We introduce some of the most recent in this poster.

Flambé is a machine learning experimentation framework built to accelerate the entire research life cycle. Flambé's main objective is to provide a unified interface for prototyping models, running experiments containing complex pipelines, monitoring those experiments in real-time, reporting results, and deploying a final model for inference. Flambé achieves both flexibility and simplicity by allowing users to write custom code but instantly include that code as a component in a larger system which is represented by a concise configuration file format. We demonstrate the application of the framework through a cutting-edge multistage use case: fine-tuning and distillation of a state of the art pretrained language model used for text classification.